Brain-cognition relationships and treatment outcome in treatment-resistant late-life depression

Abstract

Older adults with treatment-resistant depression are at significant risk for cognitive impairment. However, the relationship between treatment response and cognitive function in this population is not well-established. We examined neural correlates of executive and memory function, and their relationship with remission to pharmacological treatment in a biomarker study embedded within a clinical trial for late-life treatment-resistant depression. Among participants who completed neuroimaging, better cognition was associated with lower connectivity between components of the default mode and the frontoparietal networks and within the frontoparietal network. Using diffusion imaging data, lower tract integrity in a distributed set of tracts was associated with poorer processing speed. Additionally, gray matter structure was positively associated with cognition. Greater education was also associated with better cognition. Ongoing treatment resistance was predicted by poorer cognition and gray matter structure in cross-validated regularized logistic models. These results identify distinct cross-sectional associations between neural circuits and cognitive function in treatment-resistant late-life depression, and build a predictive model for antidepressant medication response.

Late-life depression is hard to treat and presents a risk for dementia. Here, the authors identify neural correlates of cognitive function in late-life depression, and pinpoint clinical symptoms, gray matter loss and worse cognition as robust baseline markers of non-remission to antidepressants.

Introduction

Treatment-resistant depression is a chronic, debilitating disorder defined by the persistence of depressive symptoms despite two adequate trials of antidepressant medications. Treatment-resistance is particularly challenging in late-life depression (LLD), as further antidepressant trials are unlikely to produce remission and chronicity of the depression is likely^1–3. In addition, treatment resistant LLD is accompanied by high rates of cognitive impairment^4,5 and significantly increased risk of future dementia⁶, with up to 4–6 fold increases following a recent depressive episode (i.e., within 10 years)⁷. Conversely, remitted LLD patients do not show significant differences in brain structure and function from healthy controls⁸. Impairment in executive function and episodic memory is found in both LLD and in early dementia. Similarly, frontal executive and cortico-limbic circuits underpinning those key cognitive functions are also dysfunctional in both LLD and dementia⁹.

The shared circuitry underlying LLD and dementia includes hippocampal and cortico-limbic changes¹⁰ as well as fronto-executive dysfunction, potentially via frontostriatal ischemia⁹. Biomarkers common to both LLD and dementia have been identified using structural¹¹, functional, and diffusion magnetic resonance imaging (MRI). Although findings in case-control studies of LLD are highly heterogeneous^10,12,13, network mapping based approaches have localized structural differences in major depression to frontoparietal, dorsal attention and visual networks, some of which also encompass posterior parietal and medial temporal areas affected in early Alzheimer’s Disease^14,15. Further, brain-cognition studies identify lower connectivity between frontoparietal and default mode networks and worse executive function and memory in older adults with depressive symptoms¹⁶ and non-depressed older adults with varying levels of cognitive dysfunction¹⁷. Finally, disruption of axonal white matter tracts, measured using white matter hyperintensities^18–20 has been shown in both dementia and LLD^21–23. However, there is a paucity of studies investigating neural mechanisms linking cognitive impairment and depression neurobiology in LLD, and no well-powered clinical trials to date have prospectively collected a wide breadth of precision biomarkers in treatment-resistant LLD.

Here, we present an analysis of the baseline neuroimaging and cognitive data from the Optimizing Outcomes of Treatment-Resistant Depression in Older Adults – Neurocognitive and Neuroimaging Biomarkers (OPTIMUM-NEURO) study, where we investigate the neural correlates and protective factors for cognitive function in these high-risk older individuals. These participants also participated in the “OPTIMUM” clinical trial¹, which evaluated antidepressant switch and augmentation strategies. In functional connectivity analyses, we focus on large scale networks derived from the UK Biobank²⁴, and leverage state-of-the art white matter tractography to assess brain-wide tract integrity in relation to cognitive function. We hypothesized that loss of gray and white matter alongside de-segregation of the frontoparietal and cortico-limbic circuits will be associated with worse cognitive performance in cross-sectional analyses. We hypothesized white matter disruptions to be associated with worse frontal-executive function and processing speed²⁵. We also expected level of education, a known proxy for cognitive reserve and resilience²⁶, to show protective effects on brain circuits and cognition. Finally, we tested the ability of baseline imaging and cognitive data to predict acute treatment outcomes in the OPTIMUM clinical trial, expecting the addition of neuroimaging features to improve prediction performance.

Results

The sample was predominantly female, and we observed a wide range of neuropsychological performance, with over 40% of the sample assigned a neuropsychological diagnosis of MCI (Fig. 1, Table 1).

Fig. 1. Overview of study design and analyses.

We included participants with various levels of cognitive function, MRI and clinical data in our analyses (A). Across all participants, we conducted three multivariate partial least squares (PLS) regressions (B). The first PLS model tested for associations between functional connectivity (FC) and cognitive function; the second PLS model tested for associations between fractional anisotropy, a measure of white matter tract integrity, and cognitive function; the third PLS tested for associations between gray matter and cognitive function. A total of 6 cognitive domains were included. We followed up the PLS analyses by testing for associations between PLS brain and cognitive scores with years of education and brain structural reserve (C). In addition, in a separate set of analyses we used cross-validated logistic regression models to predict patients’ remission status using clinical, demographic, cognitive and neuroimaging markers (D). Area-under-the-curve and confusion matrices were used to assess and visualize model performance. The time at which “baseline“ neuropsychological and MRI assessments were completed relative to the parent OPTIMUM trial is shown in the bottom panel, with more details available in Supplementary Section S2. Abbreviations: OPTIMUM-NEURO optimizing outcomes of treatment-resistant depression in older adults, MCI mild cognitive impairment, MADRS Montgomery-Asberg Depression Rating Scale, LV latent variable, P1-PN participants 1:N, NP neuropsychology and MRI assessment. This figure was created in BioRender. Felsky, D. (2026) https://BioRender.com/l1d0307.

Table 1.

Overview of sample demographics

	OPTIMUM-NEURO sample characterization
Demographic and clinical variables of interest	Characteristics of sample with neuropsychological data		Characteristics of sample with both neuropsychological and MRI data
N	397		234
Age (Years)	68.2	(5.9)	67.6		(5.4)
Female Sex (N)	268	68%	168	72%
Race (W/AA/A/NA)	354/28/8/7		209/16/4/5
Ethnicity (H/Non-H)	375/22		228/6
Education (Years)	14.8	(2.5)	14.7		(2.7)
MADRS score	19.5	(9.0)	19.7		(9.1)
Remission rate Step 1	31.0%		32.8%
Remission rate Step 2	23.0%		25.9%
Diagnosis (NC/MCI/DEM/NA)	178/197/13/9		116/105/6/7
CDR (0/0.5/1/2/NA)	168/182/8/1/36		111/96/5/0/22
ATHF score	8.0	(2.8)	8.2		(2.8)
CIRS-G score	8.3	(4.5)	8.9		(4.2)
MoCA score	24.6	(3.0)	24.9		(4.7)

Means and standard deviations are shown. The following diagnoses were made by adjudication among neuropsychologists: CIRS-G provides a measure of total medical burden, where a score of 8 indicates roughly 4 moderate-level conditions (eg hypertension). CIRS-G provides a measure of total medical burden, where a score of 8 indicates roughly 4 moderate-level conditions (eg hypertension). An ATHF score of 8 or more indicates that participants failed at least two adequate antidepressant trials. CDR: Clinical Dementia Rating Global scores, where 0=cognitively normal, 0.5: Questionable Impairment/Mild Cognitive Impairment (MCI); 1: Mild Dementia. NA: Not answered; There were no differences between the N = 397 who completed neuropsychological testing in the context of the clinical trial, with the N = 234 who successfully completed both neuropsychological testing and MRI.

NC normal cognition, MCI mild cognitive impairment, DEM dementia, W White Caucasian, AA African American, A Asian, NA not answered, H Hispanic, Non-H non-Hispanic, MADRS Montgomery-Asberg Depression Rating Scale, ATHF antidepressant treatment history form, CIRS-G Cumulative Illness Rating Scale – Geriatrics, MoCA Montreal Cognitive Assessment.

FC patterns associated with cognitive function

A PLS regression model with two latent variables showed that functional connectivity explained 8.4% of variance in 12 cognitive tests, significantly more than expected by chance (p_PERMUTATION= 0.03, Fig. 2A). The model included two latent variables. The first latent variable (FC-PLS1) was significantly correlated with attention, immediate and delayed memory, language and executive function (Fig. 2B). We found that greater connectivity of the DMN with the FPN (e.g., between independent components (IC) 1 and 5) was robustly associated with worse cognitive performance across multiple domains. More details on IC definition can be found in the Supplementary Information. Lower connectivity of left and right FPN components with each other was also robustly associated with better cognitive performance. Higher connectivity of visuo-motor connectivity was associated with better cognitive performance. In total, 17 connectivities showed significant loadings on FC-PLS1 (Fig. 2C). Supplementary FC analyses with a precuneus seed provided a more specific brain connectivity map linked to cognitive function. A generalizability analysis of held-out data from each of the four sites showed modest generalizability in all sites except for UCLA that showed higher generalizability (Fig. 2D). Finally, we found that better cognitive performance summarized by the FC-PLS1 cognitive scores was significantly associated with years of education (p = 1.3 × 10⁻⁵) and with brain structure centile (p = 0.00016), but not with ATHF (r = −0.08, p = 0.26). The second latent variable (FC-PLS2) captured a large amount of variance in the FC data, but was not significantly associated with cognition.

Fig. 2. Associations between resting-state functional connectivity (FC) and cognitive function.

Lower FC between DMN and FPN components was significantly associated with better cognitive performance (A, permutation distribution is shown in gray and the observed value in red; one-sided permutation p = 0.03). Higher cognitive scores on FC-PLS1 were associated with better cognitive performance on a range of cognitive tests (E, B, two-tailed P_BONFERRONI < 0.05). Lower connectivity of frontoparietal and default mode network components with each other (shown in blue), and higher connectivity of visual network components (shown in red) significantly contributed to FC-PLS1 brain scores (|Z| > 3) and was associated with better cognitive performance (C). Further, we observed modest generalizability of the results in held-out samples (D). Models were trained on data from three of the four sites. We indicate held-out sites used for model testing in (D). Predicted cognitive scores were calculated as the first principal component of variation in the five cognitive tests significantly associated with PLS1 in (B) to ensure consistency across cognitive scores derived from all data splits. Years of education (F) and brain structure centile scores (H) had protective effects on cognitive function measured as the PLS latent variable cognitive scores. Education showed a small, non-significant association with latent variable brain score (G). FC-PLS functional connectivity partial least squares regression, CRB cerebellum, STR striatum.

Reduced fractional anisotropy associated with worse psychomotor function and processing speed

A PLS regression model testing for a relationship between FA in 62 tracts and cognition explained 1.6% of variance in all cognitive tests, significantly more than expected by chance (p_PERMUTATION= 0.033, Fig. 3A). FA data specifically predicted Trail Making A performance but not the performance on any other cognitive test. The model included one latent variable (WM-PLS1), which was significantly correlated with the psychomotor speed on the Trail Making A (Fig. 3B). In total, 33 tracts significantly contributed to the WM-PLS1, with lower FA in those tracts predicting lower psychomotor function and processing speed (Fig. 3D). A prediction analysis in held-out data showed moderate-to-high generalizability in all sites (Fig. 3C). Finally, we found that better cognitive performance summarized by the WM-PLS1 cognitive scores was significantly associated with years of education (p = 0.0009, Fig. 3E) and with brain structure centile (Fig. 3G, p = 0.00035), but not with resistance to antidepressant treatment (ATHF r = −0.08, p = 0.23). White matter hyperintensities were significantly associated with both WM-PLS1 scores representing white matter integrity (r = 0.397, p = 1.6 × 10⁻⁹) and with the cognitive scores (r = 0.17, p = 0.01).

Fig. 3. Associations between tract integrity measured using fractional anisotropy (FA) and cognitive function.

Higher FA in a range of tracts including corpus callosum, longitudinal tracts, superior frontal and superficial tracts was associated with better psychomotor/processing speed, as the overall PLS model explained a significantly larger amount of variance than expected by chance (A, permutation distribution is shown in gray and the observed value in red; one-sided permutation p = 0.033). Two-tailed univariate Pearson’s correlations between individual cognitive tests and tracts are shown in (B), with tracts that significantly contributed to WM-PLS1 (|Z| > 3) highlighted in black on the top of the panel. One motor and processing speed test (Trail Making A) significantly contributed to WM-PLS1 (*P_BONFERRONI < 0.05). Multivariate correlation between cognitive WM-PLS1 scores and white matter integrity WM-PLS1 scores is shown in (D). We observed good generalizability of the results for Trail Making A performance in all held-out samples (C). Education level (E) and brain structure centile scores (G) had protective effects on processing speed measured as the PLS latent variable cognitive scores. We did not observe significant associations between education and PLS brain scores (F). WM-PLS white matter partial least squares regression, ILF inferior longitudinal fasciculus, IOFF inferior occipitofrontal fasciculus, mdLF middle longitudinal fasciculus, Sup. Frontal superior frontal, EF executive function.

Regional brain structure associated with cognitive function

A PLS regression model testing for a relationship between cortical and subcortical brain structure and cognition explained 14.5% of variance in cognitive function, significantly more than expected by chance (p_PERMUTATION= 0.009, Fig. 4A). The PLS model generated three latent variables. Two of those variables (GM-PLS1 and GM-PLS3) were significantly correlated with memory performance (Fig. 4B, Bonferroni P < 0.05). GM-PLS2 was significantly correlated with the non-memory related cognitive domains. We did not observe significant associations between clinical measures of resistance to antidepressant treatment at baseline (ATHF) with PLS latent variable scores.

Fig. 4. Associations between brain structure and cognitive function.

Higher cortical thickness in the insular cortex and greater volume of the hippocampus were associated with better cognitive performance in a range of cognitive tests, as the overall PLS model explained a significantly larger amount of variance than expected by chance (A, permutation distribution is shown in gray and the observed value in red; one-sided permutation p = 0.009). Two-tailed univariate Pearson’s correlations between individual cognitive tests and latent variable scores are shown in (B, *P_BONFERRONI < 0.05), with regions that significantly contributed to GM-PLS1 and GM-PLS2 (|Z| > 3) shown in (C). Higher cortical thickness of the insular cortex and medial temporal regions (shown in orange) and normalized volumes of the bilateral hippocampus were positively associated with cognitive function, while thickness of the superior frontal gyrus was negatively associated with cognition (shown in blue). Multivariate correlation between brain structure scores and cognitive scores for GM-PLS1 and GM-PLS2 are shown in (F) and (G), respectively. We observed good generalizability of the prediction of in most hold-out samples (D, E). Specifically, leave-one-site-out cross-validation analyses were performed to predict the first principal component of variance in the cognitive tests significantly associated with brain structure in (B). Years of education had protective effects on cognitive function (H) and brain structure latent scores (I). GM-PLS: gray matter partial least squares regression.

Identifying Markers predicting treatment outcomes

Cross-validated elastic net regression models (Supplementary Information) showed varied prediction performance for remission (MADRS ≤ 10) to bupropion or aripiprazole treatment in step 1 of the OPTIMUM clinical trial, and to lithium or nortriptyline in step 2 of the clinical trial, depending on the included predictors. The OPTIMUM trial included two steps, each 10 weeks’ duration. In step 1, patients were randomly assigned 1:1:1 to a switch to bupropion or augmentation of their current antidepressant with bupropion or aripiprazole. In Step 2, patients who were ineligible for step 1, or who did not remit or otherwise benefit from their step 1 treatment, were randomly assigned 1:1 to a switch to nortriptyline or lithium augmentation. First, we found a cross-validation AUC of 0.67 when including cognitive data alongside baseline MADRS and demographics to predict remission to treatment offered in step 1 arm of the trial, compared to AUC = 0.65 for baseline MADRS and demographic model only (n = 290). In the sample with both cognitive and neuroimaging data (n = 162), the AUC for the model with baseline MADRS, demographic and neurocognitive data was 0.67 vs 0.66 for the clinical and demographic model. This increased to an AUC of 0.73 when also including 74 brain structure variables (Fig. 5A–C). When testing the best performing model including 24 cortical thickness and subcortical volume variables, memory recall, executive function and attention scores, and baseline MADRS as predictors, the out-of-sample AUC increased to 0.83 (specificity = 0.70, sensitivity = 0.80). Neither resting-state fMRI nor diffusion derivatives improved classification performance of remission in step 1 or step 2. When modeling treatment outcomes in step 2 arm of the trial, cortical thickness but not cognitive or clinical data was predictive of remission (Fig. 5–F). When testing the best performing model including 19 cortical thickness or subcortical volume variables and baseline MADRS as predictors, the out-of-sample AUC increased to 0.77. Finally, we operationalized treatment response as change in MADRS scores, and found that a PLS regression with demographic, clinical, cognitive and brain structure data achieved considerable in-sample and hold-out accuracy (Fig. 5G–J; PLS permutation P < 0.001) in predicting change in MADRS scores.

Fig. 5. Treatment outcome prediction using both cognitive and neurobiological markers.

Elastic net regression models predicting response to treatment in step 1 of the OPTIMUM trial performed best when brain structure and cognitive data were included as predictors (A, B). Cortical thickness loadings contributing to the most parsimonious model are shown in (A), whereby regions with positive loadings are highlighted in red and regions with negative loadings are highlighted in blue. Elastic net regressions predicting response to treatment in step 2 performed worse than the step 1 models; cognitive data did not improve performance in held-out data (D, E). B, E Show the area-under-the curve (AUC) for each model set, with bar charts showing mean and standard deviation of AUCs across the 100 folds randomly sampling 20 participants on each iteration for step 1 and 15 participants on each iteration in step 2. We dropped models in which no predictors survived regularization and AUC was equal to 0.5. Black bars show the model performance for the best performing model with a narrow set of predictors, wherein variable selection was not cross-validated, i.e., conducted across the whole sample. Confusion matrices with true and false positives and negatives for the best-performing models predicting remission in steps 1 and 2 are shown in (C, F), respectively. A partial least squares (PLS) regression including demographic, clinical, cognitive and all available brain structure predictors explained a significant amount of variance (39.7%) in slopes of change in MADRS scores (G), with moderate predictive accuracy in held-out data (H). Some of the significant predictors (|Z| > 3) and their correlation with the MADRS slopes are shown in (I). Example correlations between change in MADRS scores and baseline MADRS or postcentral thickness are show in (J). MADRS Montgomery-Asberg Depression Rating Scale, CT cortical thickness, FA fractional anisotropy. All p-values shown are two-tailed.

Sensitivity analyses

We ran three additional FC PLS models, testing for the brain cognition associations in run 1 and run 2 separately, and in 196 individuals with mean FD < 0.5. The results of the main analysis were consistent with the results of the sensitivity analyses (Supplementary Section 5). Consistent with the whole brain analyses of pairwise connectivity of network components, seed-based connectivity analyses of the precuneus, overlapping with the posterior-medial default mode IC1, also showed that reduced connectivity with inferior parietal and inferior frontal regions and increased connectivity with entorhinal and perientorhinal cortex was associated with better cognitive function (SI Section 4). Second, we repeated the hold-out analyses by splitting the sample according to the scanner used (GE vs Siemens Prisma). Prediction accuracy was substantially lower in held-out data of a different scanner type (SI Section 6). Third, given that the cognitive and neuroimaging visits occurred after the completion of OPTIMUM treatment for some participants (SI Section 3), it is possible that treatment has also impacted these markers. Therefore, we repeated some of the treatment outcome prediction analyses, while only including participants who completed cognitive and neuroimaging visits before the end of their treatment. The prediction results remained highly consistent, with relatively strong performance of the best model in this subsample, too (SI Section 7).

Discussion

In this study, we identify distinct neural correlates of poorer cognitive function in people with treatment-resistant LLD. As hypothesized, we found that de-segregation between frontoparietal and default mode circuits and loss of gray matter predicted worse cognitive function. Further, worse baseline cognitive function and brain structure predicted ongoing treatment resistance, with robust cross-validation performance improvements (AUC ≥ 0.73, and up to 0.83 with feature selection) obtained by adding cortical thickness predictor features.

We found that higher connectivity between components of the frontoparietal and default mode networks was associated with worse memory recall, processing speed, and executive function. Higher between-network and lower within-network connectivity, i.e., de-segregation of these brain systems, is typically found in older compared to younger adults^27–29, is linked to worse working memory²⁸ and episodic memory^27,30, and is also found in major depression³¹ and MCI³². Resting-state connectivity of the posterior default mode regions and frontoparietal regions including the dorsolateral prefrontal cortex is a biomarker for better memory performance in older adults^16,17 and posterior parietal activity is related to better memory performance³³. Recent brain stimulation studies showed that transcranial magnetic stimulation of the precuneus showed preserved cognition³⁴. Our findings support further investigation of the posterior default mode circuits as a key correlate of cognitive function and a potential intervention target for slowing cognitive impairment in LLD.

We also found that lower white matter integrity of several tracts (e.g superior longitudinal fasciculus) was specifically associated with worse Trail Making A performance. This finding is consistent with previous studies showing that reduced integrity of tracts connecting prefrontal and parietal areas, including the superior longitudinal fasciculus, is associated with worse processing speed in healthy individuals³⁵, in those with varying levels of cognitive impairment²⁵, but also in children³⁶ and young adults³⁷. Similarly, faster processing speed is associated with higher FA in tracts including corpus callosum^38,39 and average cerebral FA⁴⁰, and cortico-striatal tracts⁴¹. Unlike gray matter and functional connectivity, FA was not significantly associated with episodic memory performance in our analysis, consistent with previous studies of white matter^41,42. We also show that white matter hyperintensities, an index of cerebrovascular disease burden, were predictive of both cognitive function and FA scores, consistent with the theory that cerebrovascular pathology affecting white matter may be driving cognitive impairment⁹. Our findings further support the hypothesis that initial impairment in white matter in treatment-resistant LLD may predominantly impact processing speed rather than memory and language domains, while functional connectivity is associated with a range of cognitive tests.

Our follow-up analyses also identified several protective factors for cognitive function in LLD. First, greater education predicted better cognitive performance and higher white matter integrity. While a consistent positive effect of education on cognitive ability in late-life has solidified its role as a proxy of cognitive reserve, there is conflicting evidence for education effects on rates of cognitive decline^43–48 and Alzheimer’s Disease⁴⁹. In addition, higher brain structure centile scores also showed protective effects on cognition, consistent with previous studies showing lower brain centiles in MCI and dementia⁵⁰ and executive function in healthy older adults⁵¹. Our findings demonstrate these relationships in patients with treatment-resistant LLD, and largely align with the literature in other late-life populations. They also provide important insights into key factors protecting against cognitive impairment for this at-risk population.

In addition to identifying brain-cognition relationships, our predictive analyses advance the search for biomarkers predicting remission in LLD in several ways. Cognitive performance and gray matter integrity predicted remission to bupropion or aripiprazole treatment in the OPTIMUM trial. Patients with better cognitive function and greater rostral ACC and postcentral gyrus thickness were more likely to achieve remission. The use of comprehensive neuropsychological data (along with demographic and clinical predictors) provided a reasonable AUC of 0.67 (in both samples). However, the addition of neuroimaging in the model boosted the AUC to 0.73. These results showcase the utility of biomarkers in helping target specific medications to patients who are more likely to benefit from them^52–54. Recent work by our group⁵⁵ and others^56,57 showed that biomarkers are less generalizable when tested out-of-trial, especially when the patient populations are different in age, severity, or other clinical features. Adult MDD biomarker studies have shown functional connectivity^55–58 and task-based activation^54,59 to be important predictors of remission. In older adults with considerable variability in gray matter integrity and cognitive performance, we found gray matter structure but not brain function to predict remission, potentially suggesting distinct biomarkers in treatment-resistant LLD and adult MDD. This contribution of gray matter integrity is independent from age, as age was included in the models but was not predictive. Cross-trial generalizability studies^55,56 are needed to test whether unique biomarkers apply in different MDD patient populations. Cognitive performance predicted remission to bupropion or aripiprazole in step 1, but not to lithium or nortriptyline in step 2, suggesting that different markers may be predictive for remission to distinct drug classes. This discrepancy may be because of different treatments used in step 2, or because patients in step 2 were even more treatment-resistant than those in step 1, with lower step 2 remission rates ( ~ 20% vs ~30%) reflecting that many had already failed a step 1 treatment. The structural MRI markers predicting remission are remarkably stable over the span of a few months, suggesting this marker may have prognostic potential, which can be tested directly in future studies.

Our study has several strengths and limitations. First, we leveraged a unique, deeply-phenotyped sample of patients with treatment-resistant LLD with advanced neuroimaging data to test for multivariate brain-cognition associations. We use a coarse parcellation derived using a data-driven group independent component analysis from a large sample of older adults in the UK Biobank^24,60, in line with recent evidence showing that brain activity can be parsimoniously explained by geometrically constrained brain-wide modes of brain geometry⁶¹. We corroborate the findings from this whole-brain approach using a seed-based FC analysis. The effect size of brain-cognition relationships identified here was moderate, reflected in modest out-of-sample generalization of the predicted cognitive scores. Sensitivity analyses showed that our brain-cognition findings were only significant in the combined group, suggesting that greater heterogeneity in cognitive performance⁶² across cognitive subgroups helps identify brain-cognition associations. Further, although cognitive function and brain MRI were assessed at varying time intervals relative to the OPTIMUM treatment, our supplementary data show that brain structure and cognition were remarkably stable over a 6-month time period. It is possible that OPTIMUM treatment has impacted functional and diffusion MRI, and we found these markers did not predict remission. Both brain and cognitive data were collected at the same interval relative to the OPTIMUM treatment, supporting our cross-sectional brain cognition findings. Future work analyzing longitudinal data collected at later time points will further illuminate change in MRI and cognitive measures over time as well as prediction of treatment response over time. Such work can help further clarify issues of potentially different timescale or sensitivity to treatment of structural brain measures and cognitive outcomes.

Previous brain-cognition association studies in remitted LLD and MCI have found similar association levels between brain structure and cognitive function²⁵. While the performance of our remission prediction models was substantial (AUC > 0.75), it can be further improved and needs to be tested in cross-trial analyses to establish generalizability^55,56. The sample size available for our neuroimaging models is a limitation. Previous work has shown that even for medium sample sizes like those in our study, models may overestimate predictive performance⁶³. Following prior recommendations, we aimed to ensure sizeable test sample sets, varying between 4-fold and 8-fold cross-validation. We also found clinical models ran in the smaller subset of participants with MRI and neurocognitive data to be highly consistent with clinical models ran in the larger subset of any participants who had clinical data. Future studies including more participants from minority groups and may uncover more cross-cultural protective factors.

In conclusion, varying levels of cognitive performance in older adults with treatment-resistant LLD have distinct neural correlates, with protective effects of education and structural brain maintenance. Our findings are consistent with existing literature in healthy aging and MCI, suggesting that many brain-cognition findings such as the role of DMN and FPN in memory and executive function, and protective effects of education and intact brain structure, are critical in treatment-resistant LLD. Our predictive models are promising and should be tested prospectively as treatment-relevant biotypes (1UG3MH137353-01). Our analyses lay the foundation for ongoing prospective longitudinal analyses to determine who among patients with treatment-resistant LLD is at highest risk of neural and cognitive decline, and how that risk relates to treatment response vs. resistance.

Methods

All research complies with ethical regulations, and ethical approval was obtained from the institutional review boards of each of the five sites—Washington University in St. Louis; Columbia University; the University of California, Los Angeles; the University of Pittsburgh; and the University of Toronto. An overview of the study design is shown in Fig. 1.

Participants

All participants were enrolled in the parent “Optimizing outcomes in older adults with treatment-resistant depression” (OPTIMUM) trial, with detailed protocol, and primary outcome results recently published^1,64, and concurrently enrolled in the present OPTIMUM-NEURO study which added MRI scans and detailed neuropsychological testing. The trial was conducted in accordance with the Good Clinical Practice guidelines of the International Council for Harmonization and was governed by an independent data and safety monitoring board. Ethical approval was obtained from the institutional review boards of each of the five sites. This report does not include participants from the Columbia University site due to an ongoing pause for all human subject research in the Department of Psychiatry at that site; this pause currently precludes the analyses of data for any ongoing human subject research study. All sample sizes reported exclude participants from Columbia University. Inclusion/exclusion criteria for the clinical trial are summarized in the Supplemental section and are previously published. Written informed consent was obtained from all the patients before enrollment and they were compensated for their time. To take part in the OPTIMUM trial, patients had to be over 60 years old and have a diagnosis of current major depression according to DSM-5 criteria which persisted despite two or more trials of antidepressants of adequate dose and duration as classified by the Antidepressant Treatment History Form (ATHF) within the current episode⁶⁵. Patients who appeared to have dementia (defined by a Short Blessed Test ≥10) were excluded. Treatment resistance was determined by research staff using a PHQ-9 score of 6 or higher at baseline in spite of 2+ adequate antidepressant trials in the current depressive episode; this was later amended to PHQ-9 score of 10 or higher. In addition, patients were required to be taking one adequately dosed antidepressant. Exclusion criteria included severe neuropsychiatric and neurodegenerative conditions such as Parkinson’s Disease or schizophrenia, uncorrected sensory impairment, imminent risk for suicide, and moderate-to-severe substance or alcohol use disorder. Patients were recruited via referrals from primary care providers and psychiatrists, outreach from the trial team, automated alerts in electronic medical records and print, radio, social media, and office advertisements. Participants who had no contra-indications for MRI scanning were offered participation in the OPTIMUM-NEURO study, that was funded to evaluate the trajectories of cognitive function (focusing on memory, processing speed and executive function) and brain structural and functional decline (focusing on cortico-limbic and fronto-executive circuits). The parent trial included two steps, each 10 weeks’ duration. In step 1, patients were randomly assigned 1:1:1 to a switch to bupropion or augmentation of their current antidepressant with bupropion or aripiprazole. In Step 2, patients who were ineligible for step 1, or who did not remit or otherwise benefit from their step 1 treatment, were randomly assigned 1:1 to a switch to nortriptyline or lithium augmentation.

Clinical data

We used the Montgomery-Asberg Depression Rating Scale (MADRS) to assess depression severity closest to the time of the initial MR scan and neuropsychological testing visit⁶⁶, and the change in depressive symptoms due to the treatment offered through the OPTIMUM clinical trial. Finally, we used the Cumulative Illness Rating Scale—Geriatrics (CIRS-G) to quantify general disease burden⁶⁷.

Cognitive data

Participants’ scores were normalized against benchmark data provided by Delis-Kaplan Executive Function System (DKEFS) and Repeatable Battery for the Assessment of Neuropsychological Status (RBANS, Supplementary Information). Cognitive domain scores for immediate and delayed memory, verbal fluency and attention and language from RBANS and an executive domain score from DKEFS were included because these cognitive processes are most affected by depression, normal aging and dementia. In the full sample with neuropsychological data (n = 397), only 3–5% of scores on each test were missing and were imputed using the mean to ensure that we could run multivariate analyses on the full sample. Cognitive status (no cognitive disorder, MCI, and mild dementia diagnoses) was determined via adjudication following DSM-5 and NIA-AA 2011 criteria with a multi-disciplinary team of clinical research staff, neuropsychologists, and psychiatrists based on the cognitive assessment scores and a wide range of other relevant clinical factors, including medical history, occupational and educational attainment, etc.

MRI data acquisition

We acquired high quality T1-weighted, diffusion MRI (dMRI) and resting-state fMRI sequences on 3T whole-body scanners using harmonized Adolescent Brain and Cognitive Development study protocols⁶⁸ across the five sites (Supplementary Table 1).

T1 structural data

We used FreeSurfer (Version 6.0.0) to derive total intracranial volume, gray, white, and CSF volumes and brain structure centiles quantifying deviations from normative data in over 100,000 individuals⁵⁰. These centiles were used as a proxy for brain maintenance⁶⁹. In addition, we utilize T1-FreeSurfer-derived total volume of white matter hyperintensities, corrected for total intracranial volume, and transformed using the square root function to ensure normality of the skewed distribution as an alternative index of cerebrovascular health⁷⁰. In region-wise analyses of brain structure, we included FreeSurfer-derived cortical thickness in the Desikan-Killiany parcellation⁷¹ and lateralized volumes of the hippocampus, amygdala and striatum corrected for total intracranial volume (aseg parcellation). Images with a total number of surface holes >380 were excluded⁷².

Resting-state functional MRI (fMRI) connectivity data

Two runs of fMRI data were pre-processed using fmriprep⁷³; the resulting minimally preprocessed images (in the NLin6 MNI space) were denoised by regressing out 24 noise components (Supplementary Information) and smoothed with a Gaussian kernel with full-width half measure of 3 mm. The first three volumes were discarded to reach steady-state equilibrium. Demeaned and normalized timeseries from the two timeseries were concatenated, with partial least square (PLS) results from individual runs available in the Supplementary Information. Participants with mean framewise displacement FD < 0.7 were kept, and sensitivity analyses with a more stringent threshold of mean FD < 0.5 are presented in the Supplementary Information.

Diffusion weighted imaging data

Diffusion image pre-processing followed previous studies⁷⁴ and included (i) brain masking (using AFNI and MRtrix3 dwi2mask), (ii) motion and eddy current correction (FSL eddy), and (iii) susceptibility distortion correction (BrainSuite BDP). We used 3D slicer to fit DTI tensors and reconstruct white matter tracts via deterministic unscented Kalman filter tractography⁷⁵ (https://github.com/SlicerDMRI). Next, we ascertained individual white matter tracts by clustering fibers and applying supervised groupwise registration to the ORG (O’Donnell Research Group) atlas^76–78. Finally, we analyze the fractional anisotropy (FA) as a measure of tract integrity. Among 73 reconstructed tracts, we excluded tracts with more than 3% unusable data and imputed missing data in the remaining 62 tracts using mean imputation.

MRI data harmonization and quality control

Batch and site artifacts can present a challenge in multi-site trials such as OPTIMUM-NEURO. The most important mitigation step is prospective harmonization, which was done here via the use of ABCD protocols at all five sites. In addition (Supplementary information), we batch normalized the data for age, sex, and site for both functional connectivity and diffusion data using ComBat⁷⁹; we also included average head motion (mean framewise displacement) in the harmonization of the fMRI data. Visual quality control of each data modality output was completed, and participant scans were excluded when anatomical segmentation of gray or white was inadequate, too much motion was present or registration between modalities was inadequate.

Statistical analyses

We used three partial least squares regressions (plsregress.m, SIMPLS algorithm⁸⁰, MATLAB R2022a) to identify latent variables capturing multivariate relationships between functional connectivity and cognitive function (FC-PLS), between fractional anisotropy and cognitive function (WM-PLS) and between gray matter structure and cognition (GM-PLS). Model significance was tested using permutation testing following previous studies^16,81 (n = 5000). The PLS permutation procedure tests whether the model with one, two or three latent variables jointly explains more variance than expected by chance. Permutation involves random row-wise reshuffling of the Y data on each permutation, and then running a PLS regression, while keeping the correlation structure of the original X and permuted Y data.

We z-scored the predictor matrix X and the outcome matrix Y (X = 211 × 210 and Y = 211 × 6 in the FC-PLS; X = 219 × 62 and Y = 219 × 10 in the WM-PLS; X = 212 × 74 and Y = 212 × 6 in the GM-PLS). PLS returns a set of latent variables that attempt to maximize the covariance between the PLS scores summarizing X and Y. PLS scores are a linear combination of the predictor variables (X) and component loadings. We used bootstrapping (n = 5000) to identify predictors that showed robust contributions to the each PLS latent variable. A threshold of |Z| > 3 was chosen to identify the most robust connectivities significantly associated with cognitive performance (see above). We correlated the latent brain scores (XS) with the cognitive tests and applied Bonferroni correction (P < 0.05) to identify cognitive tests significantly associated with each latent variable.

Robustness analysis in held-out data

To evaluate the robustness of PLS performance in each of the sites, we split our participants into four subsamples, one for each included site. We used three of these subsamples as training data and the remaining subsample as test data. We applied the PLS beta regression coefficients obtained in the training sample to the test sample and correlated the observed cognitive data with the predicted cognitive data to assess PLS performance in predicting cognitive function in held-out data⁸². Instead of keeping all cognitive tests, we created a composite cognitive variable using a principal component analysis of variables significantly associated with FC, tract integrity, and brain structure in the FC-PLS, WM-PLS, and GM-PLS respectively. All code is publicly available at https://github.com/peterzhukovsky/brain_cognition_TRD.

Prediction of treatment outcomes

We used regularized, cross-validated elastic net logistic regressions (lassoglm.m, MATLAB R2022a) to predict remission (MADRS ≤ 10) to approximately 10 weeks of acute antidepressant treatment in the parent OPTIMUM trial. Step 1 and step 2 were considered as separate studies. On each of 100 iterations, we split the data into training and test datasets, featuring 20 randomly selected participants in the test dataset for step 1 and featuring 15 randomly selected participants in the test dataset for step 2. This approximately corresponds to 8-fold and 6-fold cross-validation in the outer fold. On each iteration, we used 10-fold cross-validation to train the elastic net models in the inner fold. Area-under-the-curve (AUC) measures were then used to assess model performance in the held-out test data. Three sets of prediction models were run, each including demographic, clinical, cognitive data and one of the three imaging modalities. The clinical model included age, sex, and baseline MADRS scores. These models were run in the larger sample (n = 397) including cognitive data, clinical data, and demographic data. Then the model was re-run in the sample (n = 234) that also included neuroimaging to determine the potential improvement in the AUC when adding neuroimaging. Stringent cross-validation approach ensured that model training was completely separate from the test data. Following this iterative process, we selected the most parsimonious model that included predictors surviving regularization in over 95% of models ran. We then tested the performance of this most parsimonious model in an even 8-fold and 5-fold split of the step 1 and step 2 data, respectively, presenting confusion matrices of this model showcasing the true positives and negatives as well as false positives and negatives. Feature selection in this most parsimonious model was affected by train and test data, with cross-validation applying to the regression weights but not to the variable selection as variable selection process was done across the whole sample. In addition, we have used a PLS model predicting MADRS change (absolute difference in MADRS score between baseline and the end of last step completed) from clinical (MADRS), demographic (age, sex), cognitive and brain structure data. More information on predictive modeling can be found in the Supplementary Information.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

A.N.V., M.A.B., B.H.M., E.J.L., J.F.K., P.B., H.L., S.R., J.S.S., A.J.F., and D.M.B. designed and supervised the OPTIMUM clinical trial and OPTIMUM-NEURO biomarker study. P.Z. and A.N.V. are responsible for the analytic design of this study. P.Z., N.S., F.O., A.R., K.C., Y.A., E.W.D., and M.A.B. processed data; P.Z. led data analyses. P.Z., M.A.B., H.L., P.B., J.S.S., E.J.L., D.M.B., A.J.F., J.F.K., S.R., E.W.D., D.F., G.E.N., Y.A., N.S., F.O., A.R., K.C., B.H.M., and A.N.V. performed the research and wrote the paper.

Peer review

Peer review information

Nature Communications thanks Anouk Schrantee and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

(OPTIMUM was funded by the Patient-Centered Outcomes Research Institute; OPT-NEURO ClinicalTrials.gov number, NCT02960763, NCT05531591 and OPTIMUM Award TRD-1511-33321, while OPTIMUM-Neuro was funded by the NIMH via a collaborative R01 mechanism (Pittsburgh: MH114969; Washington University: MH114966, CAMH/Toronto: MH114970; UCLA: MH114981; Columbia: MH114980). Registration was first submitted on October 28 2016.

Data availability

The data for the OPTIMUM-Neuro study can be obtained via the NIMH Data Archive (NDAR ID 2851, https://nda.nih.gov/study.html?id=2851). Data from the parent OPTIMUM clinical trial are available upon request from the authors, and the data sharing statement was included as part of the publication of the original trial (Lenze EJ, et al. N Engl J Med. DOI: 10.1056/NEJMoa2204462). Source data will be accessible on the NIH portal as per NIH guidelines.

Code availability

We share all code used in the manuscript on GitHub and Zenodo: https://github.com/peterzhukovsky/brain_cognition_TRD/ and https://doi.org/10.5281/zenodo.20802145.

Competing interests

P.Z. was funded by the CIHR postdoctoral fellowship. A.N.V. currently receives funding from CIHR, the NIMH, the National Sciences and Engineering Research Council (NSERC), the CAMH Foundation, and the University of Toronto. BHM holds and receives support from the Labatt Family Chair in Biology of Depression in Late-Life Adults at the University of Toronto. He currently receives or has received, within the past 5 years, research support from Brain Canada, the Canadian Institutes of Health Research, the CAMH Foundation, the Patient-Centered Outcomes Research Institute (PCORI), the US National Institute of Health (NIH), Capital Solution Design LLC (software used in a study funded by the CAMH Foundation), and HAPPYneuron (software used in a study funded by Brain Canada). Within the past 5 years, B.H.M. has also received research support from Eli Lilly (medications for an NIH-funded clinical trial) and Pfizer (medications for an NIH-funded clinical trial). He has been an unpaid consultant to Myriad Neuroscience. AJF has received grant support from the U.S. National Institutes of Health, Patient-Centered Outcomes Research Institute, Canadian Institutes of Health Research, Brain Canada, Ontario Brain Institute, Alzheimer’s Association, AGE-WELL, the Canadian Foundation for Healthcare Improvement, and the University of Toronto. DMB receives research support from CIHR, NIMH, PCORI, Brain Canada and the Temerty Family through the CAMH Foundation and the Campbell Family Research Institute. He received research support and in-kind equipment support for an investigator-initiated study from Brainsway Ltd. He was the site principal investigator for three sponsor-initiated studies for Brainsway Ltd. He also received in-kind equipment support from Magventure for two investigator-initiated studies. He received medication supplies for an investigator-initiated trial from Indivior. He is a scientific advisor for Sooma Medical. He is the Co-Chair of the Clinical Standards Committee of the Clinical TMS Society (unpaid). The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76842-4.