Machine learning models of plasma proteomic data predict mood in chronic stroke and tie it to aberrant peripheral immune responses

Abstract

Post-stroke depression is common, long-lasting and associated with severe morbidity and death, but mechanisms are not well-understood. We used a broad proteomics panel and developed a machine learning algorithm to determine whether plasma protein data can predict mood in people with chronic stroke, and to identify proteins and pathways associated with mood. We used Olink to measure 1,196 plasma proteins in 85 participants aged 25 and older who were between 5 months and 9 years after ischemic stroke. Mood was assessed with the Stroke Impact Scale mood questionnaire (SIS3). Machine learning multivariable regression models were constructed to estimate SIS3 using proteomics data, age, and time since stroke. We also dichotomized participants into better mood (SIS3 > 63) or worse mood (SIS3 ≤ 63) and analyzed candidate proteins. Machine learning models verified that there is indeed a relationship between plasma proteomic data and mood in chronic stroke, with the most accurate prediction of mood occurring when we add age and time since stroke. At the individual protein level, no single protein or set of proteins predicts mood. But by using univariate analyses of the proteins most highly associated with mood we produced a model of chronic post-stroke depression. We utilized the fact that this list contained many proteins that are also implicated in major depression. Also, over 80% of immune proteins that correlate with mood were higher with worse mood, implicating a broadly overactive immune system in chronic post-stroke depression. Finally, we used a comprehensive literature review of major depression and acute post-stroke depression. We propose that in chronic post-stroke depression there is over-activation of the immune response that then triggers changes in serotonin activity and neuronal plasticity leading to depressed mood.

1.1. Introduction

Stroke is a major cause of disability worldwide,¹ and depression is a common and debilitating complication.² Approximately 30–50% of survivors experience depression after stroke, with the highest incidence in the first year.^3,4 Two years after stroke, the prevalence of depression remains higher than in the general population, at about 20%.⁵ Stroke survivors with depression have a worse quality of life and higher mortality rates than those without depression.²

Despite the persistence and severity of mood problems in chronic stroke, post-stroke depression is not well understood at a molecular level. Notably, stroke survivors experience depression at higher rates than others with similar functional disability, suggesting a stroke-specific contribution to mood.⁶ Prior history of depression is a risk factor within the first 6 months post-stroke.^7,8 Female gender, advanced age, and lesion location have also been inconsistently associated with post-stroke depression.^7,8 In those more than 6 months after stroke (chronic post-stroke depression) cognitive decline, fatigue, loss of independence and inability to work are most commonly associated with depression.³ Because depression risk remains elevated for years, it is plausible that this chronic post-stroke depression is not merely a reaction to acute stress but rather may be a biological process initiated by a stroke that chronically increases depression risk.⁴

A promising candidate for a common mechanism between post-stroke and major depression is inflammation, given that neuroinflammation occurs in the stroke scar for decades,^9,10 and that chronic peripheral inflammation can produce neuroinflammation.¹¹ In major depression without stroke, aberrant immune activation has been implicated in downstream alterations in monoamine activity, glutamate activity, and growth factor levels that induce depressive symptoms.^12,13 Demonstration of peripheral inflammation in major depression has not been straightforward, however. There are reports of increased blood levels of C-reactive protein, tumor necrosis factor (TNF)-α, Interleukin (IL)-1β and IL-6 in people with major depression, but none have been consistently observed.^14–16 In chronic post-stroke depression, large studies with broad panels of plasma biomarkers are lacking.^14–16

To address this, we utilized a proteomic approach. We measured a broad panel of 1,196 proteins in plasma samples from 85 chronic ischemic stroke survivors using Olink technology. The proximity extension assay employed by Olink produces highly specific and reproducible measures in small plasma volumes, and it has been successfully utilized in post-stroke populations.^17,18 We selected this broad protein array to avoid missing novel pathways that may be associated with mood, while simultaneously assessing evidence of abnormalities in candidate pathways in the blood. To establish whether there is a statistical relationship between plasma proteins and mood, we first utilized a machine learning algorithm to ask whether the plasma protein levels could be used to predict mood scores using either the proteomics data alone, or with the addition of age and time since stroke. We incorporated the proteomics data into multivariable regression models along with relevant clinical features, and asked whether our model could predict mood. Next, we used univariate analyses to identify individual proteins associated with mood and also tested whether candidate proteins differed based on mood in our cohort. Finally, we integrated our data in the context of existing literature to construct a mechanistic model of chronic post-stroke depression.

2.1. Methods:

2.1.1. Study Participants

Participants from the StrokeCog study¹⁹ who completed the Stroke Impact Scale (SIS) survey were included if their study assessment was completed 5 months or more after their stroke and they were over 25 years old. This study was approved by the Stanford University Institutional Review Board. Inclusion criteria for the parent study are 1) history of ischemic stroke confirmed by MRI or CT scan, 2) ability to return for annual follow-up visits, and 3) fluency in English. Exclusion criteria are 1) severe aphasia or other cognitive impairment severe enough to preclude completion of the neuropsychological battery, 2) life expectancy of less than one year, and 3) pre-existing dementia or other neurological, psychiatric, or other conditions (e.g., vision impairment or epilepsy) that would impact the assessment of neurologic and/or cognitive outcomes. History of depression was not an exclusion criterion.

2.1.2. Clinical Data & Questionnaires

We collected demographic data, and all participants underwent a 60-minute comprehensive neuropsychological battery assessing multiple cognitive domains.¹⁹ In addition, the Stroke Impact Scale survey (SIS version 3.0), a well-validated and commonly used assessment in stroke survivors, was used to assess functional abilities and symptoms after stroke.²⁰ It includes 8 subsections: strength, cognition, mood and emotional functioning, communication, activities of daily living, mobility, hand functioning, and participation in meaningful activities. The third subsection of the Stroke Impact Scale, SIS3, assesses mood and emotional functioning.²⁰ SIS domain scores were transformed to a normalized scale ranging between 0 and 100. We prorated scores from the SIS self-reported questionnaires for any missing values as long as more than 50% of scale items were answered. For all questionnaires, scores from individual questions were reversed if needed so that lower scores consistently represent more problematic symptoms.²¹ For SIS3, this means that worse mood scores are lower than better mood scores.

2.1.3. Plasma Collection & Protein Analysis

Blood was obtained at the time of the study assessment, between 5 months and 9 years following stroke onset (Figure 1A). Blood was drawn into sterile BD Vacutainer tubes with 15% K3 EDTA solution (Thermo Fisher Scientific) and placed on ice, then centrifuged within 30 minutes at 3,000 rpm for 10 minutes at 4°C. The plasma supernatant was removed, aliquoted, and stored at −80°C prior to use for proteomic analysis by Olink (Boston, MA). We used all available Olink proteomics panels, which yielded 1,196 proteins per sample.

Figure 1: Clinical workflow and feature correlations.

A) Eighty-five people with acute ischemic stroke participated in this study. Plasma was obtained between 5 months and 9 years following stroke onset, and at the same time participants completed the Stroke Impact Scale (SIS). Plasma samples were processed by Olink Proteomics, and subsequent data was analyzed. B) Clinical data were incorporated into a correlation network to visualize the relationship between them, where edges reflect a Spearman correlation coefficient of ≥ 0.1. The network is visualized using a layout calculated by the tSNE algorithm. Teal circles represent the -log10(p value) of clinical features that are significantly associated with the proteomics data, with larger circles representing more significant associations. SIS surveys contained the following missing values: 13 SIS1, 1 SIS5, 1 SIS6, 10 SIS7, 1 SIS8. C) A correlation network was generated to display the relationships between individual proteins and SIS3. Each node represents an individual protein, and node size reflects the correlation coefficient between a given protein and SIS3 score. Node color represents the -log10 p-value of the relationship with SIS3 using a Spearman test, and edges reflect a correlation coefficient of ≥ 0.1 between nodes. SIS=Stroke Impact Scale.

2.1.4. Machine learning prediction of SIS3 using plasma proteomics and clinical data

A multinomial logistic regression model was used to estimate SIS3 scores based on proteomics and clinical data. Olink data is reported as normalized protein expression (NPX), a unit in Log2 scale that allows comparison of relative protein concentration between samples. There were 186 proteins excluded from the model due to NPX values of some samples below the detection limits of the Olink platform. In total, 85 participant samples and 1,011 proteins were included in the machine learning modeling of SIS3. The ability of the machine learning algorithm to predict SIS3 was evaluated by testing the correlation of the predictions (model output) and the actual measured SIS3 scores using Pearson tests. In addition, we incorporated age, time since stroke, National Institutes of Health Stroke Score (NIHSS) score, stroke size, and sex as features in the machine learning models. Scikit-learn libraries in Python 2.7.18 and Julia 1.5.3 were used to develop the models.

2.1.5. Univariate analysis

For univariate analyses, the Spearman correlation coefficient was utilized to measure the correlation between the concentration of each individual protein and SIS3 score. Given that the Spearman correlation analysis is able to compensate for missing values, unlike machine learning models, all 1,196 proteins measured by Olink were included in this analysis. Participants were also dichotomized by SIS3 score into better mood (SIS3 > 63), or worse mood (SIS3 ≤ 63). This cutoff of 63 has been used in prior research²¹ as it represents the summed median score between “problematic symptoms” (i.e., ratings from 1–3) and “not problematic symptoms” (i.e., ratings of 4–5) on individual SIS3 items. To visualize dichotomized data, NPX values were inverse log transformed and normalized to the average within the better mood (SIS3 > 63) group (p’ = 2^p/AVG(N)) where N and AVG stands for the group with better mood and average function for statistical tests, respectively. Dichotomized data were then processed using GraphPad Prism 9 software. Statistical outliers were removed utilizing the ROUT method (Q = 0.1%), and to statistically compare the two groups for candidate proteins a two-tailed Student’s t test was used. Statistical significance was defined as p < 0.05, and the Benjamini-Hochberg method was used to correct for multiple comparisons.

2.2. Source Code and Data availability

Proteomics measurements, de-identified clinical data, and source code for independent reproduction of results are available through https://nalab.stanford.edu/strokecog/

3.1. Results:

3.1.1. Study subjects and workflow

A total of 85 participants from the StrokeCog cohort were included in this study. These were sequentially enrolled participants who completed the Stroke Impact Scale (SIS) at least 5 months after their stroke (Figure 1A). Demographic characteristics are summarized in Table 1, with subjects dichotomized into better mood (SIS3 > 63) and worse mood (SIS3 ≤ 63).²¹ Participants’ median age was 65 years old, and 41.2% were female. Although the lesion location was relatively consistent between groups, the worse mood group had a significantly larger stroke size, and higher NIHSS scores at the initial hospital stay. Median time from stroke was 341 days, and significantly longer (824 days) in the worse mood group. Additionally, participants in the worse mood group were more likely to be taking antidepressants, have a diagnosis of depression, and to use tobacco. Other clinical characteristics such as the presence of hypertension and hyperlipidemia were similar between groups.

Table 1: Demographic information of study participants.

The presence of heart disease includes a heart attack, atrial fibrillation, coronary angioplasty, bypass surgery, a pacemaker, congestive heart failure, angina, heart valve repair, or other cardiovascular disease. Moderate physical activity is defined as at least 1 hour per week of physical activity. Antidepressant medications are Bupropion, Citalopram, Duloxetine, Escitalopram, Fluoxetine, Mirtazapine, Paroxetine, Sertraline, Venlafaxine, Trazodone and Nortriptyline. Statistics represent the results of a Mann-Whitney test of the better mood vs. worse mood groups for continuous variables, or a Fisher’s exact test for dichotomous variables.

	Overall Population [median; interquartile range]	Better Mood (SIS3 > 63) N=68	Worse Mood (SIS3 ≤ 63) N=17
Female Sex	41.2%	38.2%	52.9%
Age (Years)	65; 55–71	6; 58–72.75	64; 49.5–70
Race/ethnicity  White  American Indian or Alaskan  Asian  Black or African  American  Native Hawaiian or Pacific Islander  Other/Unknown Baseline Clinical  Characteristics (%)	65.9% 1.2% 17.6% 0% 1.2% 14.1%	70.6% 0% 16.2% 0% 0% 13.2%	47.1% 5.9% 23.5% 0% 5.9% 17.6%
Heart Disease  BMI  Diabetes	37.6% 26.5; 23.6–30.4 20%	41.2% 26.0; 23.6–30.3 16.2%	23.5% 27.3; 23.3–33.7 35.3%
Hypertension  Hyperlipidemia  Current tobacco use  Current drug use  Current alcohol use  Diagnosis of depression	51.8% 34.1% 3.5% 1.2% 61.2% 21.2%	50% 35.3% 0% 1.5% 66.2% 16.2%	58.8% 29.4% 17.6%** 0% 41.2% 41.2%*
Lesion location (%)  Left  Right  Bilateral	40% 54.1% 5.9%	41.2% 51.5% 7.3%	35.3% 64.7% 0%
Time since Stroke (Days)	341; 225.5–1169	309; 215–1057	824; 326.5–1592*
Stroke Size (mL)	4.9; 0.53–35.5	; 0.33–19.3	39.4; 8.2–51.9**
NIHSS Score (initial hospital stay)	4; 1–8	3; 1–6	8; 5–13***
MoCA Score	25.5; 22–28	26; 23–28	24; 19.5–28
Antidepressant Use (% using)	23.5%	13.2%	64.7%****
Lifestyle Risk Factors
Education Years  Marital Status (% Married)	16; 16–18 67%	16; 16–18 70.6%	16; 16–18 52.9%
Moderate Physical Activity	81.2%	83.8%	64.7%

P<0.05 is considered statistically significant.

^*P<0.05;

^**P<0.01;

^***P<0.001;

^****P<0.0001

3.1.2. Relationships between clinical features and plasma proteomics data

We first utilized correlation networks to visualize the relationships between individual clinical features and proteomics data. To retain power, all data were inputted as continuous variables. The correlation network in Figure 1B revealed that individual subsections of the self-reported SIS surveys clustered closely together, while other clinical variables such as sex and age formed a second cluster. When we incorporated the plasma proteomics data into the correlation matrix of clinical features (Figure 1B), SIS3 (Mood) significantly correlated with plasma proteomics data. Age, SIS5 (Activities of daily living) and SIS6 (Mobility) also significantly correlated with plasma protein levels. This demonstrated a mathematical relationship between the proteomic features and mood, implying a biological relationship.

We next visualized the interconnectivity of individual plasma protein levels (proteomic features) and their association with SIS3 scores in a correlation network (Figure 1C). Proteins that are significantly correlated with SIS3 cluster closely together in the correlation network.

3.1.3. Machine learning modeling of SIS3

Encouraged by these associations between proteomics data and SIS3 score on our exploratory analyses, we next developed a machine learning model to statistically define these relationships. First, we asked whether we could train a machine learning algorithm to use the proteomics data to estimate, or predict, SIS3 score. To utilize the best of our data and make the model approximately unbiased, we used a leave-one-out cross-validation approach to train the model. In each iteration of this method, the model is blind to one subject’s data. Data from all other subjects is used to train the model, and that “left out” subject’s data is used for testing the model. This is repeated until the algorithm establishes whether a model can be constructed to predict the outcome (in this case SIS3 score). This leave-one-out strategy of cross-validation is widely accepted in the field of machine learning to ensure that models are correct and not overfitting the data.²²

Model estimations visualized with the Pearson test statistics demonstrated that plasma proteomics data alone have a significant predictive power towards SIS3 (Figure 2; Supplementary Table 2). Specifically, the RMSE for the model including all proteins was 20.86 and the Pearson statistic was 0.24 (-log10p-value: 1.54; Supplementary Table 2). Given the robust predictive power of the machine learning model towards data from previously unseen subjects and the reported test statistics, we concluded that our model is not over-fitting the data. Additionally, given the stronger relationship between proteomic data and SIS3 score utilizing the Pearson test (compared to the Spearman test), we concluded that the relationship between these variables is likely linear. We next asked whether incorporating individual clinical variables in the model training would improve the power of the model to estimate SIS3 scores. The model’s accuracy was enhanced by individually adding both time since the stroke (RMSE 20.13, Pearson statistic 0.28 with -log10p-value 2.08, Spearman statistic 0.24 with -log10p-value 1.54) and age (RMSE 19.93, Pearson statistic 0.3 with -log10p-value 2.23, Spearman statistic 0.3 with -log10p-value: 2.28) to the proteomics data (Supplementary Table 2). NIHSS, stroke size, and sex did not improve model accuracy (data not shown). The combination of age, time since stroke, and plasma proteomics data generated the most statistically significant machine learning model with an RMSE of 18.81 and Pearson statistic of 0.41 with a -log10p-value 4.08 (Spearman statistic 0.36 with a -log10p-value 3.08; Figure 2; Supplementary Table 2).

Figure 2: Machine learning models of SIS3 score.

A machine learning model was generated to determine whether Olink proteomics data (1,011 total proteins) can be used to estimate SIS3. Additional clinical features of subject age, and time since stroke were incorporated to improve the model. Bar height shows the -log10(p value) (Pearson test) of the correlation between the model outcome and measured SIS3. The dashed line represents p = 0.05.

We next asked whether we could identify a smaller subset of proteins with strong power to estimate SIS3 score. Commonalities in the function of proteins within this subset could shed light on underlying biological mechanisms of post-stroke depression. To ascertain this limited group of proteins, a univariate Spearman correlation analysis was utilized to identify individual proteins from the Olink dataset (1,196 total proteins) that were statistically correlated with SIS3 scores. A total of 202 proteins significantly correlated with SIS3 (p < 0.05; Supplemental Table 1). However, 22 of these proteins could not be utilized in the machine-learning model because some samples included NPX values below the Olink detection limits. We incorporated only the remaining 180 correlated proteins into our established machine learning algorithm, to see if these alone were contributing to the model. The converse was true – limiting the proteomics input to our multivariable model vastly reduced its power to estimate SIS3 score (RMSE 23.28, Pearson statistic −0.02 with -log10p-value 0.07, Spearman statistic 0.02 with -log10p-value 0.07; Figure 3). Thus, the accuracy of the machine learning model depends on more than these correlated proteins.

Figure 3: Comparison of the predictive models of post-stroke depression using limited sets of proteins.

The predictive power of the model for the estimation of SIS3 is highest when using data from all 1,011 proteins, along with the clinical features of participant age and time since stroke. The statistical power is reduced when only the subset of 180 proteins that are significantly correlated with SIS3 (p < 0.05) are incorporated in the model. Bar height shows the -log10(p value) (Pearson test) of the correlation between the model outcome and measured SIS3. The dashed line represents p = 0.05.

3.1.4. Investigating the Biological Pathways Associated with Post-stroke Mood

To investigate what the correlated plasma proteins tell us with regards to underlying biological mechanisms, we examined univariate Spearman correlations between individual protein levels and SIS3 scores. No individual proteins pass false discovery rate correction. As with our limited machine learning model above (Figure 3), this suggests that the statistical relationship between plasma protein levels and mood is based on many broad interactions and not due to strong effects of a limited number of proteins. We therefore used unadjusted p < 0.05 in the univariate analysis to identify 202 proteins that correlate with mood. We also designated a smaller subset of proteins as most highly correlated with SIS3 score. For this limited subgroup, we chose a cutoff of Spearman ρ > │0.3│ and unadjusted p < 0.05. We used these proteins to understand how protein levels may relate to the biological mechanism.

We first placed the 202 proteins that correlate with mood into loose functional groupings based on their known biology (Figure 4). Groups were immune proteins, integrins, growth factors, synaptic function proteins, serotonin activity-related proteins, and cell death and stress-related functional groupings, in addition to a minority that do not fall into these categories. Interestingly, most associated proteins were elevated rather than decreased in people with worse mood. Also, and excitingly, a high proportion (90/202 or 45%) were proteins that have been previously implicated in major depression, likely providing a link to the underlying mechanisms of chronic post-stroke depression.

Figure 4. Biological groupings of proteins modulated in people with chronic post-stroke depression.

Proteins that are significantly correlated (p < 0.05) with SIS3 score based on a Spearman correlation analysis are organized by their biological function, denoted in black. Proteins in blue are negatively correlated with SIS3 score, which means that expression is higher in people who report worse mood. Proteins in green are positively correlated with SIS3 score, and expression is lower with worse mood. A black arrowhead denotes proteins that have been previously associated with depression with or without stroke, based on a literature review, and a red star denotes proteins that are most highly correlated with SIS3 (ρ > │0.3│).

Another striking finding was that 80% (55 of 66) of correlated immune proteins were higher in the plasma of people with worse mood (Figure 4). Moreover, some of the 11 immune proteins that are reduced with worse mood have anti-inflammatory activities, such as Hematopoietic prostaglandin-D synthase (HPGDS)²³ or CD244.²⁴ In addition, most of the correlated proteins involved in synaptic function (7/11 proteins), growth factors (11/14 proteins), and serotonin activity-related proteins (2/3) followed this pattern of increased plasma levels in people with worse mood. Furthermore, all of the correlated proteins that are involved in mammalian target of rapamycin (mTOR) signaling, and all but 1 protein involved in cell death or the cell stress response have higher plasma levels with worse mood after stroke. In contrast, proteins categorized as integrins are almost all positively correlated with SIS3 score (10/12), suggesting that this group is unique.

Next, we examined candidate proteins that have previously been associated with major depression or are associated with pathways implicated in depression. Here we dichotomized participants to see if candidate proteins were different in chronic stroke survivors with better vs. worse mood. We first looked at three pro-inflammatory cytokines that have been extensively studied as potential plasma markers of major depression in cohorts without concurrent stroke: Interleukin (IL)-1β, tumor necrosis factor (TNF)-α, and IL-6.^2,14,15 None of these was highly correlated to SIS3 score when treated as a continuous variable in the Spearman correlation analysis (Supplementary Table 1). After dichotomization, there was also no significant difference in levels of IL-1β (p = 0.0830; Figure 5A), or TNF-α (p = 0.5287; Figure 5B). In contrast, IL-6 plasma levels were significantly elevated in subjects with worse mood after stroke (p = 0.0325; Figure 5C).

Figure 5. Normalized plasma expression of proteins of interest.

Participants were dichotomized by SIS3 score into better mood (SIS3 > 63) or worse mood (SIS3 ≤ 63). Plasma levels of the proinflammatory cytokines A) IL-1β, B) TNF-α, C) IL-6, or other immune-modulating proteins D) HPGDS, E) TRIM5 and F) IL-1RA are expressed as fold change relative to protein levels in people with better mood. Additionally, proteins which modulate serotonin activity G) QDPR, H) KYNU or I) ITGAV, or proteins involved in neuronal plasticity: J) 4E-BP1, K) GDNF and L) NRP1 are also displayed as median +/− interquartile range. Statistically significant differences are indicated as *p<0.05; **p<0.01; ***p<0.001; ****p<0.0001 based on results from a Student’s t test.

We also used Spearman correlation values to identify 3 additional immune proteins that may be better markers of immune dysfunction in people with depressed mood chronically post-stroke. HPGDS, tripartite motif containing 5 (TRIM5), and IL-1 receptor antagonist (IL-1RA) all significantly correlated with SIS3 (Supplementary Table 1) and have immune-modulatory activities. HPGDS protein levels are significantly reduced in people with worse mood post-stroke (p < 0.0001; Figure 5D). In contrast, both TRIM5 (p < 0.0011; Figure 5E) and IL-1RA (p = 0.0394; Figure 5F) are significantly increased in people with worse mood chronically after stroke.

Loss of serotonin is implicated in major depression, so we examined candidate proteins involved in homeostasis of the upstream metabolite tryptophan. Quinoid dihydropteridine reductase (QDPR) produces an essential cofactor for the rate-limiting enzyme in serotonin synthesis (tryptophan hydroxylase), and kynureninase (KYNU) is involved in the catabolism of tryptophan to nicotinamide. As continuous variables, both of these proteins are elevated with worse mood (Supplementary Table 1). After dichotomization, QDPR expression was significantly higher in people with worse mood (p = 0.0172; Figure 5G), although there was no difference in KYNU levels (p = 0.9320; Figure 5H). In contrast, integrin (ITG) αV is positively correlated with SIS3 score and thus expression is reduced with worse mood as a continuous variable (Supplementary Table 1). After dichotomization, ITGαV expression is significantly reduced in people with worse mood chronically after stroke (p < 0.0001; Figure 5I). ITGαV is the primary binding partner of ITGβ3, and this protein dimer negatively regulates serotonin neurotransmission.^25–27 Overall, this suggests that alterations in the serotonin pathway may contribute to depressive symptoms post-stroke.

Finally, loss of growth factor production may result in loss of mTOR pathway activity, and subsequent loss of neurogenesis and neuronal plasticity-related proteins which may contribute to depressive symptoms.^28,29 Eukaryotic translation initiation factor 4E-binding protein 1 (4E-BP1), a protein which suppresses neuronal translation via mTOR activity, was higher in people with worse mood in our correlation analysis (Figure 4; Supplemental Table 1). After dichotomization, expression of 4E-BP1 is increased in people with worse mood chronically post-stroke (p = 0.0034; Figure 5J). Protein levels of glial cell derived neurotrophic factor (GDNF), a neurotrophic growth factor, did not correlate with SIS3 score when analyzed as a continuous variable, however protein levels are significantly increased in people with worse mood post-stroke after dichotomization (p = 0.0343; Figure 5K). Neuropilin 1 (NRP1), a membrane-bound protein that interacts with semaphorin and plexin family members to initiate dendritic spine retraction, was decreased (positively correlated with SIS3 score; Supplementary Table 1) with worse mood after stroke. Two of these plexin family proteins also significantly correlated with SIS3 (Figure 4). Plexin domain-containing protein 1 (PLXDC1) was positively correlated with SIS3 (decreased with worse mood), while Plexin A4 (PLXNA4) was negatively correlated with SIS3 (increased with worse mood; Supplementary Table 1), although their biological roles in neuronal plasticity are not well defined. After dichotomization, protein expression of NRP1 was significantly lower in people with worse mood after stroke (p = 0.0132; Figure 5L). Overall, our data suggests that increases in inflammation, cell death, mTOR pathway activity and dysregulation of neuronal plasticity is associated with mood in people with chronic stroke.

4.1. Discussion:

We used proteomics to investigate mechanisms underlying chronic post-stroke depression. The Olink platform analyzed a comprehensive panel of 1,196 plasma proteins. We found that plasma proteomic data predicts mood in chronic stroke when inputted into a machine learning model. Mood estimates from our machine learning algorithm were improved by incorporating participant age and time since stroke. Independent univariate analyses identified individual proteins that most highly correlate with mood. These 202 proteins were organized into functional groups, and a high proportion were immune-related proteins.

In addition to demonstrating a statistical association between the entirety of the plasma proteomic data and mood, we explored potential mechanisms by looking at proteins whose individual levels were associated with mood. We observed a strong relationship between the peripheral immune system and post-stroke depression. Moreover, the vast majority of immune proteins that correlate with mood are elevated with worse mood, and of those that are reduced, several have anti-inflammatory activities.^23,24 Many immune proteins that correlate with post-stroke mood have been also been implicated in major depression, suggesting common mechanisms between major depression and post-stroke depression. Interestingly, prior studies in depression both with and without stroke have assessed inflammatory proteins such as C-reactive protein (CRP), interleukin (IL)-1β, IL-1 Receptor antagonist, IL-6 and tumor necrosis factor (TNF)-α, and found that they are not reproducibly increased in either type of depression. The general inability of the field to replicate cytokine-related findings has led some researchers to speculate that overall immune activation may be the best predictor of depressive symptoms after stroke, rather than individual protein levels.¹³ This is consistent with our observations - many inflammatory proteins are increased with worse mood in chronic stroke, and the data is most consistent with a general pro-inflammatory state associated with worse mood.

Also consistent with this hypothesis is that HPGDS, which synthesizes the anti-inflammatory prostaglandin D₂ in the blood, is lower with worse mood. Although this pathway has not been investigated in post-stroke depression, this is consistent with studies of major depression.^32,33 Prostaglandin D₂ is low in the blood of people with major depression.³³ It modulates responses and contributes to the resolution of inflammation.^23,34 It also has excellent blood-brain barrier penetration,³⁵ and brain activity has been linked to brain-derived neurotrophic factor expression.³⁶ Interestingly, loss of one receptor, CRTH2, renders mice resistant to stress-induced depression.³² Thus, loss of prostaglandin D₂ production is a promising mechanism by which changes in blood proteins might contribute to a loss of neuronal plasticity in post-stroke depression.³⁶

Our finding of increased TRIM5 with post-stroke depression is consistent with previous reports of pro-inflammatory activation of the NF-κB pathway in the brain in major depression, where rodent studies have linked it to depressive symptoms.^12,37 In contrast, increased IL-1RA may suggest that the IL-1 signaling pathway, although regulated via NF-κB activity, is down-regulated in chronic post-stroke depression. Further evaluation of this pathway is necessary to better understand its complexities, and potential role in post-stroke depression.

We propose a comprehensive model of chronic post-stroke depression mechanisms (Figure 6) that is based on our findings here and consistent with prior mechanistic literature in animal models, and portions of previous models in major depression.^38,39 It posits that aberrant immune activation in the brain and the blood persist in chronic stroke and contribute to worse mood via downstream loss of serotonin neurotransmitter activity and reduced synaptic plasticity in the brain.

Figure 6. Proposed model of chronic post-stroke depression.

Brain inflammation, particularly the Interleukin (IL)-1β and NF-κB pathways, mediates stress-induced depressive behavior in mice,^37,38 supporting the concept that inflammation may directly cause depression, represented by the pink arrow in Figure 6. Systemic inflammation can also induce depression, for example in cancer patients treated with systemic interferon (IFN)-α.³⁹ Notably, acute cytokine or lipopolysaccharide administration to rodents also induces sickness behavior including anhedonia, fever, decreased social interaction, and loss of nesting behavior.²⁹ More generally, people with an exacerbated inflammatory response during chronic illness develop depression at higher rates than the general population, and cytokine levels are associated with risk of depression in heart disease, cancer, and pregnancy.^40,41 Similarly, stroke induces a chronic immune response in both the brain and periphery which persists, even after perceived recovery.^9,10,42 Additionally, stroke survivors often have a higher burden of cardiovascular risk factors such as hypertension, obesity, diabetes, or old age, which also exacerbate the peripheral inflammatory state.^43–45 Given these connections, we posit that the chronic inflammatory response triggered after stroke may be an initiating event leading to the chronically increased risk of depression.

Our model also incorporates changes in tryptophan levels that influence serotonin levels and activity. Serotonin neurotransmission is dysregulated in major depression and many antidepressants target this pathway.⁴⁶ We report novel evidence here of reduced serotonin pathway activity with worse mood in chronic stroke via kynureninase (KYNU) and QDPR protein changes. Evidence that inflammatory cytokines can reduce serotonin neurotransmission (purple arrows, Figure 6) comes from rodent models where IL-1β, IFN-γ and TNF-α increase the conversion of tryptophan to kynurenine, reducing bioavailable serotonin.^41,47

The bottom portion of our model (blue arrows, Figure 6) proposes that pro-inflammatory changes lead to a reduction in growth factor levels, and subsequent development of depressive symptoms. This has been demonstrated in rodent studies of stress-induced depressive behavior.^37,38 We observed downstream changes in cell death (increased overall) and mammalian target of rapamycin (mTOR) pathways (reduced overall) that are congruent with disrupted synaptic plasticity in depression. The unexpected elevation of GDNF with worse mood in our data is inconsistent with acute stroke studies,⁴⁸ and could mean that growth factors are more accurate as an indicator of acute post-stroke depression, or that reduced growth factor activity in the brain may not translate to plasma protein levels.

Cell death and stress response pathways were markedly elevated as mood worsened; 96% (23 of 24) of proteins in this group increased with worsening mood. Similarly, the negative regulator of the mTOR pathway, 4E-BP1, was also elevated in people dichotomized into the worse mood group. This is consistent with decreased mTOR activity in people with major depression.^49,50 Finally, the bottom right portion of our model posits that cell death and loss of mTOR activity reduces neuronal protein translation, consistent with the reduction in neuropilin 1 we observed.

To our knowledge, the relationship between plasma proteomics and chronic post-stroke depression has not been previously studied, and our finding that plasma proteomics alone can predict chronic mood after stroke is novel. A major strength of our study is the unbiased design and ability to incorporate directionality of the associations between individual proteins and mood. Earlier studies in post-stroke depression looked acutely after stroke at a smaller subset of proteins and were unable to look at protein levels and correlate them with mood.^51,52 Both previous studies used liquid chromatography/mass spectroscopy (LC/MS) to perform proteomics. One study with 44 participants, assessed 3 months after stroke, used gene set enrichment algorithms and observed downregulation of coagulation and complement cascades with depression.⁵¹ The other study with 35 depressed participants at 1 month post-stroke reported dysregulation of lipid-associated proteins.⁵² Additional proteomics studies of acute post-stroke depression with more participants are warranted to further interrogate the plasma pathways involved in acute post-stroke depression and to relate them to those we report here in persistent depression in chronic stroke survivors.

Unbiased proteomics studies have proven critical to begin understanding the mechanisms of major depression. Four independent studies have used LC/MS and identified immune-related changes, while individual studies also identified changes in growth factors, lipid metabolism, cell signal transduction and protein metabolism in drug-naïve serum of people with major depression, and oxidative stress in serum of antidepressant-resistant people.^53–56 Given the interesting parallels between proteomic studies of major depression and our data, we postulate there may be shared mechanisms.

Although our study is the first to examine proteomics in chronic post-stroke depression, future studies must replicate these findings. One limitation of this work is that SIS3 score is not a standard method for evaluating clinical depression. However, the Stroke Impact Scale is well validated in post-stroke populations, and SIS3 score is highly correlated to more established depression measures such as the Hospital Anxiety and Depression scale (HADS) in other cohorts.^20,57 Better preclinical models of chronic post-stroke depression will also be critical to test which plasma proteins are most tightly linked to the molecular mechanisms of depression in the brain, and to validate our proposed mechanistic model.

In conclusion, we report the first proteomic analysis in chronic post-stroke depression. We utilized machine learning models to demonstrate that plasma proteomics can accurately estimate mood in people with chronic stroke. Our findings represent an important initial step in unraveling the biological mechanisms of post-stroke depression and understanding this disease in the chronic phase, when stroke survivors are primarily affected. Indeed, this study is relevant to people who are most severely affected by post-stroke depression – those with depression years after stroke and who are resistant to antidepressant treatment. We identified aberrant immune responses in people with worse mood chronically after stroke, which strongly implicates chronic inflammation in the pathophysiology of this disease. In combination with established literature on major depression, we generated a model of how inflammatory changes may lead to depressive phenotypes. Future work is needed to test these proposed mechanisms and to move towards the goal of developing therapeutics for post-stroke depression.

Supplementary Material

1

Supplementary Table 1. The list of 202 total proteins that are significantly correlated with SIS3 (p < 0.05), and their statistical relationship with SIS3 based on a Spearman correlation analysis.

2

Supplementary Table 2. Model parameters from the multivariable machine learning models in Figures 2&3.

Highlights:

• Plasma proteomic data can be used to predict mood in people with chronic stroke

• Aberrant immune responses are associated with worse mood in chronic stroke

• We infer a model of how inflammatory changes may lead to depressive phenotypes

Abbreviations:

4E-BP1

4E Binding Protein 1

BDNF

Brain-Derived Neurotrophic Factor

BH4

Tetrahydrobiopterin

CRP

C-Reactive Protein

GDNF

Glial Cell Derived Neurotrophic Factor

HPGDS

Hematopoietic Prostaglandin D Synthase

IFN

Interferon

IL

Interleukin

IL-1RA

Interleukin 1 Receptor Antagonist

ITG

Integrin

KYNU

Kyneureninase

mTOR

Mammalian Target of Rapamycin

NF-κB

Nuclear Factor Kappa B

NIHSS

National Institutes of Health Stroke Scale

NPX

Normalized Protein Expression

NRP1

Neuropilin 1

PGD2

Prostaglandin D2

PLXDC1

Plexin Domain-Containing Protein 1

PLXNA4

Plexin A4

QDPR

Quinoid Dihydropteridine Reductase

SERT

Serotonin Transporter

SIS

Stroke Impact Scale

TNF

Tumor Necrosis Factor

TRIM5

Tripartite Motif Containing 5