Proteomic and Machine Learning Analysis Predicts Treatment Response Signatures in Myasthenia Gravis

Proteomic and Machine Learning Analysis Predicts Treatment Response Signatures in Myasthenia Gravis | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Proteomic and Machine Learning Analysis Predicts Treatment Response Signatures in Myasthenia Gravis Karli Faith Gilbert, Amrita K. Cheema, Henry Kaminski, Linda Kusner This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7915342/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 4 You are reading this latest preprint version Abstract Background Myasthenia gravis (MG) is a prototypical antibody-mediated autoimmune disease with variable treatment responses with a need for biomarkers to guide therapeutic decision making. Proteomic profiling, coupled with machine learning, offers a powerful approach to identify biomarkers that may predict treatment response. Methods We analyzed sera collected at entry (baseline) from participants in a phase 3 trial randomized trial comparing thymectomy plus prednisone versus prednisone alone, along with matched controls using liquid chromatography–mass spectrometry. We derived disease-specific proteomic signatures and evaluated associations between baseline proteins and 6-month clinical outcomes using multiple machine-learning approaches with internal validation. Results Baseline serum proteomes distinguished MG from controls, with pathway enrichment implicating complement activation, immunoglobulin production, and T-cell receptor signaling. Distinct protein panels predicted 6-month clinical improvement within each treatment arm. In the thymectomy-plus-prednisone group, models captured non-linear relationships of predictive proteins in contrast with the predominant additive patterns observed in the prednisone-alone group. Predictive proteins were enriched for T-cell signaling and leukocyte trafficking functions, providing insight into treatment-specific biology. Conclusions Baseline serum proteomics captures core disease characteristics of MG and predicts short-term clinical response in a treatment-specific manner. If validated in independent cohorts, these findings could enable biomarker-guided selection of thymectomy, refine risk stratification, and furnish mechanistic readouts for future MG trials and clinical care. myasthenia gravis machine learning treatment response proteomics mass spectrometry prednisone thymectomy Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Background Therapeutic decision-making for patients with autoimmune diseases is often hindered by the central challenge of the unpredictability of treatment response. Whether due to intrinsic disease heterogeneity or variable sensitivity to immunomodulatory therapies, patients with the same diagnosis can experience markedly different outcomes. The complex interplay of many factors, including sex, age, genetic background, and environmental experiences, collectively influence disease expression and therapeutic efficacy. Myasthenia gravis (MG), a prototypical antibody-mediated disorder, exemplifies this clinical variability and offers a compelling model in which to investigate the biological underpinnings of treatment response and identify biomarkers to predict therapeutic response ( 1 , 2 ). Despite an array of groundbreaking research in the field of MG, few studies have combined rigorous, long-term clinical outcome monitoring with systematic collection of linked biological specimens. A notable exception is the NIH-supported, Thymectomy Trial in Non-Thymomatous Myasthenia Gravis Patients Receiving Prednisone Therapy (MGTX) trial, which incorporated comprehensive, validated clinical assessments linked with prospectively collected biospecimens. The MGTX trial, with a primary endpoint at 3 years, established the efficacy of thymectomy plus prednisone (ETTX) compared with prednisone alone (PA) in patients aged 18 to 65 years with acetylcholine receptor (AChR) antibody-positive MG ( 3 ). Subjects in the thymectomy arm had a lower total Quantitative MG (QMG) score and required lower overall corticosteroid dosage at end of study, although no significant differences were detected at 6 months. Prednisone is a well-established therapy for MG ( 4 , 5 ), while the application of thymectomy dates back nearly 100 years, but was only unequivocally identified as effective by the MGTX trial. However, the mechanisms underlying the benefit of thymectomy remain poorly understood. Considerable variability in treatment response was observed across both study arms. Broad-based -omics profiling approaches provide a powerful method to move beyond clinical observation toward molecular signatures that predict therapeutic outcomes. Optimization of treatment regimens in patients with MG can be supported by the identification of biomarkers predictive of treatment outcomes ( 6 , 7 ), a task increasingly enabled by machine learning approaches applied to clinical features and multi-omics data. Studies applying machine learning to investigate biospecimens from MG patients are limited and have used publicly available microarray expression profiles ( 8 – 10 ), while others have generated primary proteomic ( 11 ), metabolomic ( 12 ), and transcriptomic data ( 13 ) from MG cohorts. Here, we applied liquid chromatography–mass spectrometry (LC-MS) to sera collected at study entry (baseline) from 86 subjects of MGTX trial, of whom 41 were assigned to PA and 45 to ETTX. We first assessed for differences between MG subjects and controls and then aimed to identify serum proteins at baseline predictive of 6-month QMG improvement in treatment groups. Although subjects who had undergone thymectomy demonstrated a greater average reduction in QMG scores compared to prednisone-only subjects, about 30 percent of subjects did not achieve a well-established criterion of a clinically meaningful reduction of 3 points on the QMG score in the first 6 months of the study. We chose this early time point because of the greater uniformity of treatment early in the study based on uniform prednisone treatment protocol, an absence of rescue therapies, and expectation of thymectomy providing an early effect on the autoimmune pathology of MG. We made no prior assumptions regarding the data set and applied several machine learning (ML) models for predictors of treatment response. Methods Study Population Serum samples were obtained from participants enrolled in the MGTX clinical trial (NS42685). Specimens had been stored at -80 o C at NCI-Fredrick and then transferred to the biospecimen bank at George Washington University. Samples were collected at the baseline time point for each participant and analyzed using LC-MS. Clinical outcome measurements were collected at baseline and six months using the QMG assessment. All individuals included in this analysis had confirmed diagnosis of MG with detectable acetylcholine-receptor autoantibodies and no evidence of thymoma ( 3 ). Control serum samples were collected at George Washington University with the exclusion criteria of autoimmune disease and vaccination in the past month. All participants were older than 18 years and provided written informed consent. Patient and control characteristics are summarized in Supplemental Table 1. The NINDS funded the trial and assembled an independent Data Safety Monitoring Board. Sites received local institutional review board (IRB)/ethics committee approvals, and each patient provided written informed consent before study entry including provision of serum samples. All specimens were deidentified. The George Washington University IRB provided additional review and approved use of specimens. Mass Spectrometry Protein Analysis Shotgun proteomics of MG patient sera was performed as previously described ( 14 , 15 ). Individual samples (20 µL) and a pooled quality control (5 µL of each sample) were prepared using ENRICH-iST Kit (PreOmics®) before injection into the mass spectrometer nanoElute® 2 instrument (Bruker Scientific). Peptide fractions from the pooled quality control (9 technical replicates) were obtained using the Pierce™ High pH Reversed-Phase Peptide Fractionation Kit. A data dependent acquisition (DDA) spectral library from the peptide fractions was generated in Fragpipe and DIA_SpectLib_quant workflow. Individual samples were quantified with DIA-NN (data independent acquisition – neural network) with a maximum false discovery rate of 1%. Data Cleaning and Imputation The DIA quantification values were first filtered in Fragpipe ( 16 ) to remove contaminant proteins, reverse sequences, and proteins identified by a single peptide. The data were then converted to a log2 scale and separated by group (MG vs Control). Missing values were imputed using a Perseus-type Missing Not at Random (MNAR) approach, where values were randomly drawn from a left-shifted Gaussian distribution. This distribution was centered 1.8 standard deviations below the mean of the observed values within each group, with a standard deviation width of 0.3, to simulate low-abundance proteins below the detection limit. Analysis of Myasthenia Gravis versus Control Subjects The R package limma was used to determine the differentially expressed proteins (DEPs) between MG and healthy control (HC) samples ( 17 , 18 ). DEPs were defined as proteins that surpassed a significance cutoff of a |log 2 fold change| >1 and an adjusted p-value less than 0.05 (Benjamini-Hochberg method). Pathway enrichment analyses were performed using over representation analysis (ORA) and Gene Set Enrichment Analysis (GSEA) with the R packages clusterProfiler and enrichPlot ( 19 , 20 ). Feature Selection We applied the Boruta feature selection algorithm to identify all relevant variables for predictive detection of treatment predictive proteins. The approach enables optimization of a low sample/feature ratio, to reduce the number of proteins, and thus noise, in our predictive analyses. This allowed for enhanced performance in downstream assessments and reduced the likelihood of a classification algorithm from overfitting a model ( 21 ). Boruta The Boruta algorithm ( 22 ) was employed in R using the Boruta package with default settings, maxRuns = 300, and results recorded from 100 iterations, to obtain a subset of proteins predictive of clinical improvement in thymectomy plus prednisone (ETTX) and prednisone alone (PA) subjects separately. In each iteration, Boruta creates shadow features that are added to the dataset, runs a single random forest model on the dataset and compares the importance scores of the real variables to the shadow features. Features whose importance scores were significantly higher than the highest importance score observed among the random noise variables were retained as relevant predictors. The initial subsets of proteins were then subjected to a second set of 100 iterations of Boruta feature selection to obtain a set of proteins for PA and ETTX patients for use in nested cross validation of eight ML models. Nested Cross Validation Nested cross-validation (nCV) was employed to separate model training from testing, ensuring unbiased performance estimates and reducing the risk of overfitting in clinical outcome predictions. We separately applied a 5-fold nCV to the selected features for ETTX and PA subjects. The inner loop performed hyperparameter tuning via grid search (Supplemental Table 2) using the caret R package with its default 10-fold CV to determine the most accurate training parameters for the outer loop. The data were split across the outer folds such that each patient was tested once. Classification predictions and performance metrics were pooled or averaged across outer folds to generate unbiased estimate. nCV Models To develop robust and generalizable predictive models, we employed a diverse panel of supervised ML algorithms within the nCV framework: probabilistic classifiers (Naïve Bayes, Logistic Regression), regularized linear models (LASSO, Elastic Net), tree-based approaches (Classification Tree, Random Forest), and ensemble boosting methods (AdaBoost, XGBoost). Each model captured different aspects of data structure, which allowed us to comprehensively evaluate predictive features across varying assumptions and model complexities. This broad approach was chosen to minimize bias from any single modeling strategy and to better characterize proteomic expression patterns relevant to predicting clinical improvement. Probabilistic Classifiers We implemented Naïve Bayes and logistic regression as representative probabilistic classifiers. Naïve Bayes, which assumes independence among features, was applied using the naiveBayes function from the e1071 R package ( 23 ). Logistic regression, which models the log-odds of the binary outcome as a linear combination of predictors, was performed using the base R glm function with a binomial family. Both models are deterministic and reproducible, in contrast to stochastic tree-based and ensemble methods. Regularized Linear Models LASSO (Least Absolute Shrinkage and Selection Operator) is a linear regression technique that uses L1 regularization ( α = 1) to shrink the coefficients of less important features to zero. Similarly, elastic net is a regularized regression that balances L1 and L2 regularization (0 < α < 1). We performed LASSO and Elastic Net using the glmnet R package ( 24 , 25 ). Using the built-in cross-validation, proteins with non-zero coefficients were recorded from two regularization levels: λ min, the value of λ that minimized the cross-validation error, and λ 1se, the largest λ within one standard error of λ min. These correspond to models that favor predictive accuracy and model simplicity, respectively. Decision Trees Deterministic decision trees were made using the classification and regression tree algorithms (CART) from the recursive partitioning and regression tree R package rpart ( 26 ). For each split at a node, the relative impurity reduction was calculated as the impurity at the parent node minus the total impurity of the left and right child. A named numeric vector for each tree provided the protein names of those considered important by the algorithm. Random Forests The random forests algorithm was performed using the R package randomForest ( 27 ), which ranks features based on how much they contribute to decreasing the variance in the model, also referred to as the increase in node purity. The importance scores calculated for each protein are a normalized calculation of the contribution of each protein to the reduction of impurity in the model. Ensemble Boosting Adaptive boosting (AdaBoost) and extreme gradient boosting (XGBoost) were used as ensemble learning methods, which work to improve predictive performance by sequentially training models, each focused on correcting the errors of the previous. AdaBoost was implemented using the ada function from the JOUSBoost R package ( 28 ), while XGBoost was applied using the xgboost package ( 29 ). Both methods iteratively reweigh training observations to emphasize harder-to-classify cases. A key distinction is that AdaBoost used simple decision stumps as base learners, whereas XGBoost built full decision trees, allowing it to capture more complex patterns. A complete list of R packages and versions used can be found in Supplementary Table 3. Statistical analyses were performed in R v4.4.1 and GraphPad Prism 10.3.0. The repository with all R source code used to run nCV is available on GitHub at https://github.com/drkgil/NestCV . Pathway and Tissue Source Analysis We performed GeneAgent analysis ( 30 ) to assess biological pathways identified for proteins found to be predictive for outcomes for PA and ETTX groups. Results Proteomic profiling reveals distinctions between MG patients and controls. Eighty-six sera at study entrance to the MGTX clinical trial and 37 healthy controls were used (demographic details in Supplemental Table 1). While efforts were made to match the groups, the average age of subjects differed from controls (Control: 43.46 years; MG: 35.91 years; Mann-Whitney U = 1090, p = 0.0053); there was no difference in gender distribution. We identified proteomic signatures differentiating MG patients from controls. Liquid chromatography mass spectrometry analysis identified 1,213 proteins, from which a dataset of 1,210 were used to identify 398 differentially expressed proteins (DEPs). Among these, 23 proteins were significantly downregulated in MG compared to controls, while 375 proteins were significantly upregulated, highlighting a substantial alteration in the proteomic landscape associated with MG (Fig. 1 A). Principal component analysis (PCA) of the full proteomic dataset further revealed a clear separation between MG and control samples, emphasizing the robustness of these differences (Fig. 1 B). Among the MG patients, 77% were on prednisone at the time of study enrollment at an average dose of 30.7 mg ± 14.1 (Supplemental Table 1). We found that the use of prednisone at the time of enrollment was not a driver of the PCA separation from controls. Hierarchical clustering of the 398 DEPs further supported this distinction, with heatmap visualization clearly separating MG patients from controls, without a pattern according to prednisone use (Fig. 1 C). To explore the biological processes associated with the 398 DEPs identified in MG, we performed ORA using the Gene Ontology Biological Processes (GO-BP) pathway database. The analysis revealed numerous significantly enriched pathways, with the top 25 pathways included those related to actin filament organization and organelle organization (Fig. 2 A). ORA using the KEGG database mirrored the GO-BP results, with enrichment of terms related to actin organization and cytoskeletal dynamics (Fig. 2 B). Several of these top KEGG pathways corresponded with those typically enriched in MG patients compared to healthy controls: “leukocyte transendothelial migration”, “B cell receptor signaling pathway”, and Fc receptor pathways. To test for coordinated expression patterns that could be informative of changes between MG subjects and controls that may have escaped the threshold of significance set for DEPs, we also performed GSEA utilizing the GO-BP database. Proteins were ranked according to the B-statistic obtained from the limma analysis of Fig. 1 . While many of the top pathways overlapped with those identified by ORA(Fig. 2 C), a few significant pathways emerged. Notably, enrichment scores for immunoglobulin production (p = 7.5e-6) and complement activation (p = 0.0086) were significantly downregulated in MG patients compared to controls, while the T cell receptor signaling pathway was significantly upregulated (p = 0.03) (Fig. 2 D). Distinct proteomic signatures emerge when prednisone and thymectomy treatment groups are analyzed separately We next sought to determine which proteins could be a predictive marker of clinical improvement at 6 months based on a greater than three point improvement of the QMG score, which has been established as a clinically meaningful difference ( 31 ). The patient characteristics at baseline did not differ between the treatment groups (Supplemental Table 4). PCA of MG subject proteomics did not reveal a distinction between patients that improved compared to those that did not improve (Fig. 3 A); however, four proteins were significantly reduced in patients that showed improvement (Fig. 3 B). Similarly, the same PCA labeled by treatment group indicated a lack of clustering (Fig. 3 C), despite DEPs present between ETTX and PA treated patients (Fig. 3 D). This finding is not surprising given that the clinical trial did not show statistically significant differences between groups until after 6 months. Next, we analyzed baseline serum protein profiles separately for each treatment group. Within the ETTX and PA treatment groups, comparisons between patients who improved and those who did not revealed 16 DEPs in ETTX patients (Fig. 3 E) and 14 DEPs in PA patients (Fig. 3 F). Given the greater number of DEPs identified within individual treatment groups compared to the combined MG subjects, we maintained this separation in the data analysis. Feature selection highlights proteins predictive of improvement and distinct between treatment groups Considering the positive response of patients to thymectomy, we hypothesized that proteins predictive of clinical improvement would differ between treatment groups. The extent and nature of these differences was initially unclear as potential group-level differences were masked by the abundance of non-significant proteins, as seen in the PCA plots (Fig. 3 A, C). To address this, we independently applied Boruta feature selection within each treatment group (Fig. 4 ), identifying preliminary sets of 45 proteins in ETTX patients and 38 proteins in PA patients. A second round of feature selection was then applied to these subsets to refine the list to proteins strongly associated with improvement. This two-step selection process reduced dimensional noise, yielding 17 and 7 key proteins for the ETTX and PA groups, respectively (Table 1 ). At each step of the feature selection, we verified that the sets of features were specific to its respective treatment group by performing PCA with mismatching the patient group and predictive proteins (Supplemental Fig. 2). These group-specific protein sets were used in all subsequent analyses to test their performance across eight different machine learning algorithms in a nested cross validation design. Table 1 Baseline proteins predictive of clinical improvement in PA and ETTX treated MG patients. MG Group Protein Improved Not Improved p-value PA CCL16 7.05 ± 0.32 8.13 ± 0.39 0.04 CILP2 12.08 ± 0.12 11.61 ± 0.15 0.017 GALNT1 9.30 ± 0.30 10.27 ± 0.12 0.0046 IGHV3-43 10.73 ± 0.22 11.30 ± 0.16 0.04 RPS3 8.42 ± 0.35 7.05 ± 0.47 0.025 RPS3A 8.61 ± 0.42 7.01 ± 0.17 0.0012 VAPB 11.46 ± 0.23 11.95 ± 0.10 0.057 ETTX ARHGAP9 7.34 ± 0.34 8.35 ± 0.46 0.086 CCN5 9.47 ± 0.42 8.33 ± 0.78 0.21 CD55 6.34 ± 0.22 7.73 ± 0.49 0.019 CHAD 13.06 ± 0.15 12.38 ± 0.21 0.015 CHRDL1 9.19 ± 0.31 7.96 ± 0.51 0.051 FBLN5 10.79 ± 0.22 9.86 ± 0.36 0.038 HS1BP3 7.61 ± 0.27 8.83 ± 0.44 0.027 HSP90AA1 9.19 ± 0.42 10.70 ± 0.32 0.0068 IGKV1-5 19.85 ± 0.11 19.61 ± 0.19 0.30 MMP3 12.87 ± 0.21 13.76 ± 0.31 0.026 PCMT1 8.49 ± 0.34 7.88 ± 0.67 0.42 RAB14 10.18 ± 0.36 8.96 ± 0.63 0.11 RDX 10.34 ± 0.24 10.70 ± 0.39 0.45 RNASET2 11.03 ± 0.07 10.81 ± 0.16 0.21 SELL 14.01 ± 0.09 13.72 ± 0.17 0.14 SPP2 15.20 ± 0.13 15.76 ± 0.14 0.0059 SVEP1 11.06 ± 0.12 10.02 ± 0.26 0.002 Values reported as log 2 abundances (mean ± SEM). P-values are reported from students two-tailed t-tests between improved and not improved groups. Random Forest identifies seven proteins highly predictive of response to prednisone treatment We applied a five-fold nCV framework utilizing eight machine learning models, which decreases the potential for bias in the selected panels, on the relative abundances of 7 proteins, CCL16, CILP2, GALNT1, IGHV3-43, RPS3, RPS3A, and VAPB, expressed in PA patients at baseline. nCV is particularly well-suited for smaller datasets, as it helps prevent overfitting by ensuring each sample is independently tested from the data used to train the model (Fig. 4 ). Each model performed relatively well at classifying PA patients as improved or not, except for the classification tree algorithm. Several models showed similar performance, with receiver operating characteristic (ROC) curves and average area under the ROC curve (AUC) closely overlapping (Fig. 5 A). However, differences between models were more evident when considering the Matthews correlation coefficient (MCC) (Table 2 ), a robust metric that accounts for all confusion matrix categories. MCC scores range from − 1 to 1, indicative of total disagreement and total agreement between predicted and actual classes, respectively (Chicco and Jurman, 2020). Together, the AUC and MCC scores highlight the random forest algorithm as the best-performing classifier for PA patients in the nested cross-validation. Table 2 Performance metrics across 5 outer folds of PA patients. Metric NB Logistic LASSO EN CT RF Ada XGB AUC 0.87 ± 0.07 0.82 ± 0.07 0.87 ± 0.03 0.87 ± 0.03 0.64 ± 0.09 0.91 ± 0.06 0.74 ± 0.08 0.82 ± 0.09 AUCPR 0.95 ± 0.03 0.92 ± 0.03 0.94 ± 0.02 0.94 ± 0.02 0.75 ± 0.07 0.95 ± 0.03 0.92 ± 0.03 0.90 ± 0.06 Prec 0.90 ± 0.07 0.79 ± 0.09 0.80 ± 0.09 0.79 ± 0.07 0.71 ± 0.10 0.86 ± 0.07 0.83 ± 0.07 0.77 ± 0.08 Recall 0.68 ± 0.10 0.68 ± 0.09 0.80 ± 0.06 0.89 ± 0.05 0.72 ± 0.06 0.92 ± 0.05 0.77 ± 0.08 0.84 ± 0.09 Acc 0.76 ± 0.05 0.68 ± 0.03 0.73 ± 0.05 0.78 ± 0.05 0.64 ± 0.09 0.85 ± 0.06 0.76 ± 0.08 0.76 ± 0.07 Spec 0.88 ± 0.07 0.70 ± 0.18 0.65 ± 0.19 0.62 ± 0.12 0.47 ± 0.18 0.73 ± 0.11 0.72 ± 0.13 0.62 ± 0.12 F 1 0.76 ± 0.07 0.71 ± 0.06 0.78 ± 0.04 0.83 ± 0.04 0.71 ± 0.07 0.88 ± 0.06 0.80 ± 0.07 0.80 ± 0.08 MCC 0.56 ± 0.09 0.42 ± 0.10 0.52 ± 0.10 0.52 ± 0.10 0.17 ± 0.20 0.68 ± 0.12 0.47 ± 0.17 0.47 ± 0.14 Metrics reported as mean ± SEM. The best performing AUC and MCC metrics are in bold and the runner up in italics. NB, Naïve Bayes; EN, elastic net; CT, classification tree; RF, random forest; Ada, adaptive boosting; XGB, extreme gradient boosting; AUC, area under the ROC curve; AUCPR, area under the precision-recall curve; Prec, precision; Acc, accuracy; Spec, specificity; F 1 , F-score (harmonic mean of precision and recall); MCC, Matthew’s correlation coefficient. The random forest algorithm produced the highest AUC 0.91 ± 0.06, which was significantly better than the logistic regression (Z = -2.48, CI -0.317 to -0.037, p = 0.013), extreme gradient boosting (Z = 2.89, CI = 0.043 to 0.23, p = 0.0039), and trended towards being better than adaptive boosting (Z = 1.86, CI = -0.006 to 0.216, p = 0.063) (Fig. 5 B). Across the outer folds of the nCV, the RF model correctly classified 35/41 patients (Fig. 5 C). The mean decrease in accuracy was used to determine which of the 7 proteins were most important to the RF model in accurately predicting patient improvement outcome. The protein VAPB had the highest average importance score, followed by RPS3A , and CILP2 (Fig. 5 D). Seventeen proteins predict treatment response to thymectomy Similarly, we applied five-fold nCV to the 17 proteins predictive of improvement outcome in ETTX patients: ARHGAP9, CCN5, CD55, CHAD, CHRDL1, FBLN5, HS1BP3, HSP90AA1, IGKV1-5, MMP3, PCMT1, RAB14, RDX, RNASET2, SELL, SPP2, and SVEP1. The ROC curve analysis indicated that the adaboost and XGBoost models performed best, followed closely by LASSO, Elastic Net, Random Forest, and Naïve Bayes (Fig. 6 A, Table 3 ). The XGBoost model achieved the highest AUC (0.99 ± 0.01), performing significantly better than Logistic Regression (Z = -2.67, CI = -0.37 to -0.06, p = 0.008) and Classification Tree algorithms (Z = -5.0, CI = -0.68 to -0.30, p = 5.7e-7), and trended towards significance over Naïve Bayes (Z = -1.80, CI = -0.34 to 0.15, p = 0.072) and Random Forest algorithms (Z = -1.35, CI = -0.23 to 0.04, p = 0.18) (Fig. 6 B). Across all five outer folds of the nCV, XGBoost correctly classified 41/45 ETTX patients (Fig. 6 C). Table 3 Performance metrics across 5 outer folds of ETTX patients. Metric NB Logistic LASSO EN CT RF Ada XGB AUC 0.83 ± 0.05 0.66 ± 0.13 0.90 ± 0.08 0.90 ± 0.08 0.56 ± 0.08 0.90 ± 0.06 0.95 ± 0.05 0.99 ± 0.01 AUCPR 0.92 ± 0.05 0.89 ± 0.05 0.96 ± 0.03 0.96 ± 0.03 0.78 ± 0.05 0.96 ± 0.02 1.00 ± 0.00 1.00 ± 0.00 Prec 0.80 ± 0.07 0.84 ± 0.09 0.90 ± 0.05 0.90 ± 0.05 0.76 ± 0.07 0.79 ± 0.10 0.95 ± 0.03 0.88 ± 0.06 Recall 0.88 ± 0.07 0.76 ± 0.08 0.92 ± 0.05 0.92 ± 0.05 0.70 ± 0.11 0.97 ± 0.02 1.00 ± 0.00 1.00 ± 0.00 Acc 0.76 ± 0.06 0.71 ± 0.08 0.89 ± 0.04 0.89 ± 0.04 0.58 ± 0.05 0.78 ± 0.08 0.96 ± 0.03 0.91 ± 0.04 Spec 0.41 ± 0.19 0.58 ± 0.19 0.79 ± 0.10 0.79 ± 0.10 0.42 ± 0.19 0.44 ± 0.23 0.83 ± 0.11 0.75 ± 0.10 F1 0.83 ± 0.05 0.78 ± 0.06 0.91 ± 0.04 0.91 ± 0.04 0.69 ± 0.04 0.85 ± 0.06 0.97 ± 0.02 0.93 ± 0.04 MCC 0.40 ± 0.24 0.37 ± 0.19 0.73 ± 0.08 0.73 ± 0.08 0.08 ± 0.17 0.66 ± 0.20 0.88 ± 0.07 0.81 ± 0.08 Metrics reported as mean ± SEM. The best AUC and MCC metrics are in bold and the runner ups are in italics. SHAP (SHapley Additive exPlanations) values were calculated to explain the XGBoost model. These values correspond to how much the protein contributes to the prediction in each individual; the greater the distribution of a protein’s SHAP values, the greater its influence on the prediction of the model. Since the model performs a binary classification for each patient, positive SHAP values correspond to an increased probability of being classified as ‘improved’; conversely, negative values correspond to an increased probability of being classified ‘not improved’ (Fig. 6 D). In ETTX patients, the XGBoost model identified SVEP1, SPP2, CHAD, and RAB14 as the top important proteins for predicting clinical improvement (Fig. 6 E). GeneAgent Analysis We applied GeneAgent ( 30 ) analysis to the ETTX and PA proteins identified as predictors of clinical outcomes (Supplementary Files 6 & 7). The summary report for ETTX indicated that the extracellular matrix remodeling was the primary pathway involved in treatment response, with key indicators from cell adhesion, migration and immune response. In contrast, analysis of the seven proteins predictive of PA response revealed some association with ribosomal pathways, but no distinct summary was identified. Discussion To better understand the immunobiological underpinnings of treatment response in MG, we performed serum proteomic profiling of participants in the MGTX clinical trial and healthy controls. We identified distinct serum proteomic signatures that differentiated individuals with MG from controls, with biological pathways involving complement activation, immunoglobulin production, and T cell signaling driving these differences. Moreover, we discovered a subset of proteins that predicted clinical improvement six months later, differing between subjects treated with prednisone alone and those receiving thymectomy plus prednisone. These findings provide unique insights into the biological effects of thymectomy. Notably, application of multiple machine learning approaches revealed non-linear relationships in the identification of outcome-predictive proteins in the ETTX plus prednisone group, while independent linear (additive) patterns predominated among those receiving prednisone alone. The proteins associated with treatment response were primarily involved in T cell signaling and cell trafficking. The limitations of our work lie in those inherent to a rare disease research of small sample size, and despite our machine learning nested cross validation approach, future validation studies will be required, including those that assess change in the proteome over time. Further, the benefit of a rigorous clinical trial data set compromises the potential for immediate applicability to a real-world population with a greater diversity of patients with multiple co-morbidities. This includes patients outside the age of study entrance criteria and AChR antibody status. Myasthenia gravis driven differences from controls Our initial hypothesis was that prednisone treatment would significantly alter the serum proteome and potentially confound disease-specific differences; however, this was not observed. Only 22 DEPs were identified between subjects on prednisone and those not receiving treatment at study entry, compared to over 400 DEPs observed between MG subjects and healthy controls, regardless of prednisone use (Supplemental Fig. 1). These findings align with those of Nelke and colleagues, who reported that, in a highly heterogeneous treated cohort of MG patients, proteomic differences were greater between patients and a small control group (n = 10) than among the four patient consensus clusters ( 32 ). Despite the widespread and long-standing use of corticosteroids across a range of diseases, surprisingly little is known about their effects on gene and protein expression following chronic administration. Serum proteomic changes are presumed to result from the genomic actions of corticosteroids, which are themselves modulated by physiological adaptations, including downregulation of the glucocorticoid receptor and feedback regulation of downstream pathways ( 33 ). Moreover, glucocorticoid effects on tissue may not be reflected in the circulating proteome, as our findings suggest. The proteomic signatures differences were a function of MG. Pathway analyses revealed alterations in cytoskeletal and organelle organization. The explanation for such changes in circulation may lie in injury to the large number of neuromuscular junctions across the body. AChR antibodies induce complement-mediated injury and deposits of immunoglobulins, complement components, membrane-attack complex, and AChR are found in the synaptic cleft. Macrophage infiltration is not appreciated and therefore, these components and others could make their way into circulation. Complement components of the membrane attack complex are appreciated to be shed with tissue trauma, as would occur in MG, and these could have a pro- or anti-inflammatory effect ( 34 ). The same can be said for mechanisms of modulating effects on lymphocytes. Immune cells must migrate from blood to various lymphoid organs and other tissues. KEGG pathway analysis identified leukocyte transendothelial migration, B cell receptor signaling pathway, and Fc receptor pathways to be reduced in the MG subjects (Fig. 2 B). Similarly, GSEA evaluation of the GO-BP database was supportive of these results, but also found immunoglobulin production and complement activation were significantly downregulated in MG patients compared to controls. Complement-mediation of disease pathology of MG is well appreciated ( 35 , 36 ). T cell receptor signaling pathway was significantly upregulated (Fig. 2 D). These results may seem counterintuitive that these pathways would be reduced in an active autoimmune disorder; however, this would represent a mechanism to suppress global immune activity in the context of the autoimmune reaction. T cell receptor signaling pathway was significantly upregulated, which could be a suppressor activity. Similar pathways were identified in our transcriptional profiling analysis of thymus obtained from the MGTX study ( 37 ). Consistent with pathway analysis, the most significantly expressed DEPs were related to cytoskeletal function or directly to immune function with several overlapping. TAGLN2, PDLIM2, CNN2, and SEPTIN6 were all originally characterized by cytoskeletal association; their involvement in T cell activation and inflammatory signaling ( 38 – 40 ) make them particularly interesting proteins. CSK, or c-Src terminal kinase, is involved in T cell and B cell receptor signaling and can be found in exosomes ( 41 , 42 ). Many of the proteins identified are highly likely to be carried in exosomes, which are known to be secreted by a number of cell types in response to stress signals to promote tissue healing ( 43 ). Machine Learning Assessment To robustly capture relationships between serum proteins and clinical improvement, we applied a panel of machine learning algorithms, each with distinct pattern sensitivities. First, we used Boruta analysis to identify features that collectively contributed to the prediction of treatment response. In our study, features are proteins. Proteins that do not improve the model’s predictive performance are not selected. Importantly, proteins that may be part of the same biological pathway might not all be selected if their inclusion would not provide additional predictive value beyond what is already captured by other selected proteins. Proteins identified through feature selection were utilized in several machine learning algorithms to assess the AUC of the proteins predicting outcomes. This assumption-agnostic approach, combined with nested cross-validation, enabled head-to-head model comparison and revealed Random Forests as optimal for PA-treated patients and XGBoost for those receiving ETTX; each performed excellently with an AUC > 0.9. Random Forests and XGBoost often outperform probabilistic classifiers (Naïve Bayes, Logistic Regression) and regularized linear models (LASSO, Elastic Net) in complex biological datasets, including large proteomic data sets. This finding lies in the nonlinearity of proteomic data which is further complicated by the correlation of many proteins. XGBoost, is also tree-based ensembles that use boosting. Unlike Random Forests, where trees are built independently, boosting builds trees sequentially, with each new tree focusing on correcting the mistakes made by the previous ones. This allows the model to gradually improve its performance by learning from its own errors. XGBoost builds on this idea by adding regularization to penalize overly complex models, pruning to eliminate unhelpful splits, and parallelization to speed up training. The eight models represent a spectrum of complexity, underlying model assumptions, and learning strategies, which allowed us to systematically evaluate predictive performance across our proteomic datasets. The use of multiple ML algorithms is critical in biomarker discovery, where signals are often heterogeneous and no single method can guarantee optimal performance ( 44 , 45 ). The multi-model approach not only increases confidence that our findings were not artifacts of any one modeling assumption and provides insight into the nature of the predictive structure. Overall, these differences highlight how proteomic landscapes shape clinical response in a treatment-dependent manner, with additive effects characterizing PA subjects and nonlinear interactions driving improvement in those receiving ETTX. Previous proteomic evaluations in myasthenia gravis Our study is among the first to utilize both proteomics and ML modeling to predict short-term treatment response in patients with MG. The number of studies utilizing machine learning on biospecimen datasets, not just clinical data, from patients with MG without thymoma is sparse. This includes studies identifying biomarkers from the gut microbiome of MG using metabolomics ( 12 ) and rRNA sequencing ( 13 ), from an MG monozygotic twin microarray dataset ( 8 , 10 ), and from proteomic data from AChR-ab positive MG patients ( 11 ). A recent study by Lin et al also utilized LC-MS DIA proteomics to assess pre-surgical differences in serum between effective and ineffective thymectomy in MG; however, samples were from a cohort of patients with thymoma MG ( 46 ). Even so, they reported that the protein SELL was downregulated in treatment-responsive thymectomized patients at baseline. This finding is partially in line with our study, in which SELL was considered an important protein in predicting treatment response specifically in patients receiving a thymectomy and prednisone (ETTX patients), albeit the relative expression in ETTX patients who improved was higher than those that did not (Table 1 ). SELL (L-selectin) has been strongly implicated for its role in the adhesion and trans endothelial migration of leukocytes, especially neutrophils, before being cleared via ectodomain shedding, acting as a marker for neutrophil activation ( 47 , 48 ). Five studies used targeted (i.e., O-link) proteomics to investigate changes between MG and HC samples ( 49 – 53 ). In agreement with results reported by Bhandage et al, 2024 and Molin et al, 2017, we found elevated levels of the pro-inflammatory calcium binding protein SA100-A12 in MG patients compared to controls, which may be indicative of increased neutrophil activation and lymphocyte recruitment ( 54 , 55 ). Proteomic Signature of Prednisone Treatment Effect. DEPs at baseline were identified that were associated with improved QMG after 6 months and these were distinct for ETTX and PA groups, despite the well-matched clinical characteristics of these subjects. Seven proteins (Table 1 ) were associated with treatment response, with CCL16 and IGHV3-43 being the most clearly related to the immune response. CCL16 is a chemokine primarily expressed in the liver but also the thymus and some lymphocytes. CCL16 is appreciated as being responsible for recruitment of T cells to sites of pathology and other chemokines are well-established to be involved in MG, this has not been appreciated before for CCL16 ( 56 ). IGHV3-43 is a portion of the variable region of the heavy chain of IgG, which would be expected to be detected in serum, but why the lower level of this specific IgG heavy chain in serum would be associated with improvement is not clear. However, given that genetic variations in the immunoglobulin heavy chain locus may drive antibody response could suggest this IGHV3-43 may be of particular importance in MG ( 57 ). Results from the GeneAgent analysis similarly identified the immune response roles, and a mild predominance of ribosomal involvement by RPS3 and RPS3A. Proteomic Signature for Thymectomy plus Prednisone Treatment Effect. The 17 proteins identified by Boruta analysis as predictive of clinical improvement in MG patients treated with thymectomy plus prednisone likely reflect a coordinated biological response to immune modulation and tissue remodeling. These proteins encompass key processes including immune regulation and complement control (CD55, SELL, IGKV1-5, RNASET2), extracellular matrix and stromal reorganization (MMP3, FBLN5, CCN5), cytoskeletal dynamics and cellular trafficking (ARHGAP9, RAB14, RDX), stress response and protein repair (HSP90AA1, PCMT1), and modulation of the BMP/TGF-β axis involved in thymic epithelial architecture (CHRDL1, SPP2). Collectively, these proteins may serve as systemic indicators of thymic regression, loss of germinal center activity, suppression of autoreactive B and T cell signaling, and the resolution of inflammatory and structural abnormalities that drive disease activity. Their predictive value likely lies in their ability to capture early molecular changes that precede and enable long-term clinical benefit as observed in the MGTX trial. From a translational standpoint, the biological roles of these proteins suggest potential therapeutic targets or mechanistic entry points for understanding how thymectomy induces systemic immune reprogramming. SVEP1 had the greatest SHAP value predictive of improvement (Fig. 6 D, E). SVEP1 (Sushi, von Willebrand factor type A) is a large extracellular matrix protein that has been identified to be associated with poor outcomes in several conditions including coronary artery disease, dementia, and hypertension and with poor outcomes in heart failure, where it is strongly associated with activated T cells ( 58 – 60 ). SVEP1 is composed primarily of complement control protein (CCP, also known as Sushi) domains; whether it binds complement proteins is not known. Elevated SVEP1 levels in blood are also associated with diseases of aging, and SVEP1 has been shown to activate the mTOR pathway, a key regulator of longevity. Given autoimmune disorders increase with chronological age, these findings provide the intriguing insight that these early-onset MG patients have an advanced biological age. Heat shock protein 90α (Hsp90α) is a ubiquitously expressed heat shock protein, which interacts with close to two thousand proteins and numerous biological pathways ( 61 , 62 ). Elevations in serum are associated with greater cancer risk and vascular inflammation. Hsp90α has been shown to suppress or activate immune processes depending on specific conditions. SPP2 is produced in the liver and known to be found in circulation. It binds and activates intracellular signaling pathways of the TGF-beta superfamily cytokines. Interestingly, SPP2 shares sequence homology with osteopontin. Osteopontin gene polymorphisms are associated with treatment response to prednisone in MG ( 63 ). CD55 is a complement inhibitory protein known to be concentrated at the neuromuscular junction and protective of injury to antibody-mediated, complement-induced in experimental MG ( 64 ). While typically found on cell surfaces, CD55 exists in soluble form and in culture reduces proliferation peripheral blood monocytes and would be expected to modulate complement activity ( 65 , 66 ). CD55 has four CCP domains and therefore shares structural similarities to SVEP1. The thymus of patients with MG caused by AChR antibodies demonstrates thymic hyperplasia characterized by lymphoid follicles and germinal center formation, which was documented for the present group of patients ( 67 ). The germinal centers are a source of AChR-specific B cells and contain autoreactive T cells. Removal of the thymus reduces a source of autoreactive T and B cells but, despite its removal, clinical disease persists in a variable number of patients. The panel of proteins predictive of outcome in ETTX patients supports a fundamental difference in individual susceptibility to removal of thymus. The proteins identified and our approach to this question reveals that a single factor, such as T regulatory cell dysfunction, as appreciated in patients with MG, is unlikely to be the single answer to treatment effect of thymectomy. GeneAgent results supported critical roles of cell trafficking in these proteins; this would be consistent with the susceptibility of such pathways to treatment outcome based on the exodus of pathogenic cells from the thymus to peripheral immune organs. Clinical Impact The clinical importance of our findings lies in the identification of serum protein signatures associated with treatment response to thymectomy plus prednisone and prednisone alone, offering a potential tool for personalized treatment planning in myasthenia gravis. Currently, decisions regarding thymectomy are based on generalized criteria without reliable biomarkers to predict individual benefit ( 2 ). Our results suggest that proteomic profiling could enable clinicians to identify patients more likely to respond favorably to thymectomy, thereby reducing unnecessary surgeries in non-responders, expediting treatment decisions for those likely to benefit, and guiding others toward alternative therapies. The observed heterogeneity in treatment response is consistent with findings by Nelke et al., who identified patient subgroups with differential responsiveness, including one enriched for complement inhibition benefit ( 32 ). Further, the prednisone-only related signature would identify poor responders to the standard of care treatment for MG ( 68 ). This approach parallels precision oncology strategies, where surgical or pharmacologic interventions are tailored using molecular markers. Moreover, stratification based on this proteomic signature could inform the development of novel therapeutic targets for non-responders. Conclusions Baseline serum proteomics distinguished treatment-specific predictors of clinical response in patients with MG. PA-patients’ response was explained mainly by single protein markers, whereas ETTX-patients’ response was predicted by complex protein interactions. These findings represent an important first step toward biomarker-guided treatment in MG; with validation in independent cohorts, proteomic signatures could refine patient stratification, inform clinical trial design, and enable more personalized therapeutic decisions. Abbreviations MG: Myasthenia gravis MGTX: Thymectomy Trial in Non-Thymomatous Myasthenia Gravis Patients Receiving Prednisone Therapy AChR: acetylcholine receptor ETTX: thymectomy plus prednisone PA: prednisone alone QMG: Quantitative MG LC-MS: liquid chromatography-mass spectrometry ML: machine learning IRB: institutional review board DDA: data dependent acquisition DIA: data independent acquisition DEP: differentially expressed protein HC: healthy control ORA: over representation analysis GSEA: gene set enrichment analysis nCV: nested cross validation LASSO: least absolute shrinkage and selection operator CART: classification and regression tree algorithms AdaBoost, Ada: adaptive boosting XGBoost: extreme gradient boosting PCA: principal component analysis GO-BP: gene ontology biological processes KEGG: Kyoto Encyclopedia of Genes and Genomes SEM: standard error mean NB: Naïve Bayes EN: elastic net CT: classification tree RF: random forest AUC: area under the ROC curve ROC: receiver operating characteristic AUCPR: area under the precision-recall curve MCC: Matthew’s correlation coefficient SHAP: Shapley additive explanations CCL16: C-C motif chemokine ligand 16 CILP2: cartilage intermediate layer protein 2 GALNT1: polypeptide N-acetylgalactosaminyltransferase 1 IGHV3-43: immunoglobulin heavy variable 3-43 RPS3: ribosomal protein S3 RPS3A: ribosomal protein S3A VAPB: VAMP (vesicle-associated membrane protein)-associated protein B and C ARHGAP9: Rho GTPase activating protein 9 CCN5: cellular communication network factor 5 CD55: decay accelerating factor for complement CHAD: chondroadherin CHRDL1: chordin like 1 FBLN5: fibulin 5 HS1BP3: HCLS1 binding protein 3 HSP90AA1: heat shock protein 90 alpha family class A member 1 IGKV1-5: immunoglobulin kappa variable 1-5 MMP3: matrix metallopeptidase 3 PCMT1: protein-L-isoaspartate (D-aspartate) O-methyltransferase RAB14: member RAS oncogene family, F protein-binding protein 1 RDX: radixin RNASET2: ribonuclease T2 SELL: selectin L, lymphocyte adhesion molecule 1 SPP2: secreted phosphoprotein 2 SVEP1: sushi, von Willebrand factor A, EGF and pentraxin domain containing 1 Declarations Ethics approval and consent to participate The NINDS funded the trial and assembled an independent Data Safety Monitoring Board. Sites received local institutional review board (IRB)/ethics committee approvals, and each patient provided written informed consent before study entry including provision of serum samples. All specimens were deidentified. The George Washington University IRB provided additional review and approved these investigations. Consent for publication Not applicable Availability of data and materials The R code used to run nCV is available on GitHub at https://github.com/drkgil/NestCV. The datasets used and analyzed during the current study are available from the corresponding author on reasonable request. Competing interests All financial and non-financial competing interests must be declared in this section KG: none. AC: none. HK: receives consulting fees from Novartis, UCB Pharmaceuticals, Kite Pharmaceuticals, Canopy Immunotherapeutics, and Merck. Argenx provides an unrestricted educational grant to George Washington University. He is an unpaid consultant for Care Constitution. HK has equity interest in Mimivax, LLC. HK is principal investigator for the Rare Disease Network, MGNet supported by NIH grant U54NS115054. LK: receives consulting fees from Novartis. LK has equity interest in Mimivax LLC. LK is a co-investigator for the Rare Disease Network, MGNet supported by NIH grant U54NS115054. Funding The work was supported by the MGNet, a member of the Rare Disease Clinical Research Network Consortium (RDCRN) NIH U54 NS115054 and NIH/NINDS U01 NS042685 (MGTX). Additional support was provided by philanthropic support for the Laboratory for MG Research at George Washington University. Funding support for the Data Management and Coordinating Center (DMCC) is provided by the National Center for Advancing Translational Sciences (NCATS) and the National Institute of Neurological Disorders and Stroke (NINDS). Authors’ contributions KG: Data curation, Formal analysis, Software, Writing – original draft, Writing – review & editing. AKC: Data interpretation, Writing – review & editing. HK: Conceptualization, Data curation, Funding acquisition, Resources, Supervision, Writing – original draft, Writing – review & editing. LK: Conceptualization, Formal analysis, Funding acquisition, Methodology, Supervision, Writing – review & editing. All authors read and approved the final manuscript. Acknowledgements LC-MS performed by the Core Services at Lombardi Comprehensive Cancer Center, Georgetown University Medical Center, Washington, DC, 20057, USA. We would also like to thank the Metabolomics Shared Resource, partially supported by NIH/NCI grant P30-CA051008. References Kaminski HJ, Sikorski P, Coronel SI, Kusner LL. Myasthenia gravis: the future is here. J Clin Invest 2024 June 17;134(12):e179742. Benatar M, Cutter G, Kaminski HJ. The best and worst of times in therapy development for myasthenia gravis. Muscle Nerve. 2023;67(1):12–6. Wolfe GI, Kaminski HJ, Aban IB, Minisman G, Kuo HC, Marx A, et al. Randomized Trial of Thymectomy in Myasthenia Gravis. N Engl J Med. 2016;375(6):511–22. Alhaidar MK, Abumurad S, Soliven B, Rezania K. Current Treatment of Myasthenia Gravis. J Clin Med. 2022;11(6):1597. Verschuuren JJ, Palace J, Murai H, Tannemaat MR, Kaminski HJ, Bril V. Advances and ongoing research in the treatment of autoimmune neuromuscular junction disorders. Lancet Neurol. 2022;21(2):189–202. Fahrner LJ, Chen E, Topol E, Rajpurkar P. The generative era of medical AI. Cell 2025 July 10;188(14):3648–60. Xu C, Jackson SA. Machine learning and complex biological data. Genome Biol. 2019;20(1):76. Jin S, Yin J, Li W, Mao N. Mitochondrial autophagy-related gene signatures associated with myasthenia gravis diagnosis and immunity. Autoimmunity. 2025;58(1):2465410. Liu H, Liu G, Guo R, Li S, Chang T. Identification of Potential Key Genes for the Comorbidity of Myasthenia Gravis With Thymoma by Integrated Bioinformatics Analysis and Machine Learning. Bioinforma Biol Insights. 2024;18:11779322241281652. Zhou G, Wang S, Lin L, Lu K, Lin Z, Zhang Z, et al. Screening for immune-related biomarkers associated with myasthenia gravis and dilated cardiomyopathy based on bioinformatics analysis and machine learning. Heliyon. 2024;10(7):e28446. Schroeter CB, Nelke C, Stascheit F, Huntemann N, Preusse C, Dobelmann V et al. Inter-alpha-trypsin inhibitor heavy chain H3 is a potential biomarker for disease activity in myasthenia gravis. Acta Neuropathol (Berl). 2024 June 18;147(1):102. Chang CC, Liu TC, Lu CJ, Chiu HC, Lin WN. Explainable machine learning model for identifying key gut microbes and metabolites biomarkers associated with myasthenia gravis. Comput Struct Biotechnol J. 2024;23:1572–83. Chang CC, Liu TC, Lu CJ, Chiu HC, Lin WN. Machine learning strategy for identifying altered gut microbiomes for diagnostic screening in myasthenia gravis. Front Microbiol [Internet]. 2023 Sept 27 [cited 2025 July 15];14. Available from: https://www.frontiersin.org/journals/microbiology/articles/ 10.3389/fmicb.2023.1227300/full Bauman TA, Lee SM, Juel VC, Li Y, Gilbert K, Chen J, et al. Serum fibrinogen is not elevated in patients with myasthenia gravis. Sci Rep. 2025;15(1):13013. Cheema AK, Li Y, Ventimiglia M, Kowalczyk K, Hankins R, Bandi G, et al. Radiotherapy Induces Innate Immune Responses in Patients Treated for Prostate Cancers. Clin Cancer Res Off J Am Assoc Cancer Res. 2023;29(5):921–9. Hsiao Y, Zhang H, Li GX, Deng Y, Yu F, Valipour Kahrood H, et al. Analysis and Visualization of Quantitative Proteomics Data Using FragPipe-Analyst. J Proteome Res. 2024;23(10):4303–15. Phipson B, Lee S, Majewski IJ, Alexander WS, Smyth GK. Robust hyperparameter estimation protects against hypervariable genes and improves power to detect differential expression. Ann Appl Stat. 2016 June;10(2):946–63. Ritchie ME, Phipson B, Wu D, Hu Y, Law CW, Shi W, et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 2015;43(7):e47. Yu G. Thirteen years of clusterProfiler. Innov. 2024;5(6):100722. Yu G, Wang LG, Han Y, He QY. clusterProfiler: an R Package for Comparing Biological Themes Among Gene Clusters. OMICS J Integr Biol. 2012;16(5):284–7. Pudjihartono N, Fadason T, Kempa-Liehr AW, O’Sullivan JM. A Review of Feature Selection Methods for Machine Learning-Based Disease Risk Prediction. Front Bioinforma. 2022 June;27:2:927312. Kursa MB, Rudnicki WR. Feature Selection with the Boruta Package. J Stat Softw 2010 Sept 16;36:1–13. Meyer D, Dimitriadou E, Hornik K, Weingessel A, Leisch F, Chang CC et al. e1071: Misc Functions of the Department of Statistics, Probability Theory Group (Formerly: E1071), TU Wien [Internet]. 2024 [cited 2025 July 28]. Available from: https://cran.r-project.org/web/packages/e1071/index.html Friedman J, Tibshirani R, Hastie T. Regularization Paths for Generalized Linear Models via Coordinate Descent. J Stat Softw. 2010;33(1):1–22. Tay JK, Narasimhan B, Hastie T. Elastic Net Regularization Paths for All Generalized Linear Models. J Stat Softw. 2023;106(1):1–31. Therneau T, Atkinson B, rpart. Recursive Partitioning and Regression Trees [Internet]. 2025. Available from: https://CRAN.R-project.org/package=rpart Liaw A, Wiener M. Classification and Regression by randomForest. R News. 2002;2(3):18–22. Olson M, JOUSBoost. Implements Under/Oversampling for Probability Estimation [Internet]. 2017. Available from: https://CRAN.R-project.org/package=JOUSBoost Chen T, Guestrin C, XGBoost:. A Scalable Tree Boosting System. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining [Internet]. New York, NY, USA: Association for Computing Machinery; 2016 [cited 2025 Sept 16]. pp. 785–94. (KDD ’16). Available from: https://dl.acm.org/doi/10.1145/2939672.2939785 Wang Z, Jin Q, Wei CH, Tian S, Lai PT, Zhu Q, et al. GeneAgent: self-verification language agent for gene-set analysis using domain databases. Nat Methods. 2025;22(8):1677–85. Barohn RJ, McIntire D, Herbelin L, Wolfe GI, Nations S, Bryan WW. Reliability testing of the quantitative myasthenia gravis score. Ann N Y Acad Sci. 1998;841:769–72. Nelke C, Schroeter CB, Barman S, Stascheit F, Masanneck L, Theissen L, et al. Identification of disease phenotypes in acetylcholine receptor-antibody myasthenia gravis using proteomics-based consensus clustering. EBioMedicine. 2024 July;105:105231. Ayyar VS, Jusko WJ. Transitioning from Basic toward Systems Pharmacodynamic Models: Lessons from Corticosteroids. Pharmacol Rev. 2020;72(2):414–38. Karasu E, Eisenhardt SU, Harant J, Huber-Lang M. Extracellular Vesicles: Packages Sent With Complement. Front Immunol. 2018;9:721. Albazli K, Kaminski HJ, Howard JF. Complement Inhibitor Therapy for Myasthenia Gravis. Front Immunol. 2020;11:917. Jacob S, Howard JF. Neuromuscular junction and the complement system. Int Rev Neurobiol. 2025;182:21–42. Sengupta M, Wang BD, Lee NH, Marx A, Kusner LL, Kaminski HJ. MicroRNA and mRNA expression associated with ectopic germinal centers in thymus of myasthenia gravis. PLoS ONE. 2018;13(10):e0205464. Ivanov AI, Le HT, Naydenov NG, Rieder F. Novel Functions of the Septin Cytoskeleton: Shaping Up Tissue Inflammation and Fibrosis. Am J Pathol. 2021;191(1):40–51. Shi Y, Fan T, Yang Y, Liu J, Ouyang J, Dai J. Beyond structural domains: the emerging roles of PDLIM2 in cellular signaling and cancer progression. Front Physiol. 2025;16:1569285. Yin LM, Ulloa L, Yang YQ. Transgelin-2: Biochemical and Clinical Implications in Cancer and Asthma. Trends Biochem Sci. 2019;44(10):885–96. Hikita T, Kuwahara A, Watanabe R, Miyata M, Oneyama C. Src in endosomal membranes promotes exosome secretion and tumor progression. Sci Rep. 2019;9(1):3265. Zhu S, Wang H, Ranjan K, Zhang D. Regulation, targets and functions of CSK. Front Cell Dev Biol. 2023;11:1206539. Aoi W, Tanimura Y. Roles of Skeletal Muscle-Derived Exosomes in Organ Metabolic and Immunological Communication. Front Endocrinol. 2021;12:697204. Diaz-Uriarte R, Gómez de Lope E, Giugno R, Fröhlich H, Nazarov PV, Nepomuceno-Chamorro IA, et al. Ten quick tips for biomarker discovery and validation analyses using machine learning. PLoS Comput Biol. 2022;18(8):e1010357. Zhang X, Jonassen I, Goksøyr A. Machine Learning Approaches for Biomarker Discovery Using Gene Expression Data. In: Helder I. N, editor. Bioinformatics [Internet]. Brisbane (AU): Exon Publications; 2021 [cited 2025 July 24]. Available from: http://www.ncbi.nlm.nih.gov/books/NBK569564/ Lin X, Qi G, Liu P. Identification of potential prognostic biomarkers of thymoma with myasthenia gravis based on serum proteomics. Front Immunol. 2025;16:1580219. Ivetic A, Hoskins Green HL, Hart SJ. L-selectin: A Major Regulator of Leukocyte Adhesion, Migration and Signaling. Front Immunol. 2019;10:1068. Rahman I, Collado Sánchez A, Davies J, Rzeniewicz K, Abukscem S, Joachim J, et al. L-selectin regulates human neutrophil transendothelial migration. J Cell Sci. 2021;134(3):jcs250340. Bhandage AK, Kenina V, Huang YF, Roddate M, Kauke G, Grosmane A, et al. Serum protein biomarker profile distinguishes acetylcholine receptor antibody seropositive myasthenia gravis patients from healthy controls. iScience. 2024;27(8):110564. Bhandage AK, Punga AR. Inflammatory imbalance and activation deficits in T cells of myasthenia gravis patients revealed by proteomic profiling. Front Immunol [Internet]. 2025 Aug 4 [cited 2025 Aug 11];16. Available from: https://www.frontiersin.org/journals/immunology/articles/ 10.3389/fimmu.2025.1648020/full Molin CJ, Westerberg E, Punga AR. Profile of upregulated inflammatory proteins in sera of Myasthenia Gravis patients. Sci Rep. 2017;7:39716. Myllynen C, Sarkkinen J, Atula S, Tienari P, Kekäläinen E, Laakso SM. A skewed ratio of free light chains is more common in patients with late-onset than early-onset myasthenia gravis. Immunol Lett. 2023;260:81–8. Okuzono Y, Miyakawa S, Itou T, Sagara M, Iwata M, Ishizuchi K, et al. B-cell immune dysregulation with low soluble CD22 levels in refractory seronegative myasthenia gravis. Front Immunol. 2024;15:1382320. Bobek D, Sestan M, Mijacika L, Kovacic N, Lukic IK, Grcevic D, et al. Serum S100A12 levels in children with childhood-onset systemic lupus erythematosus, systemic juvenile arthritis, and systemic undefined recurrent fevers. Z Rheumatol. 2023;82(Suppl 1):74–9. Donato R, Cannon BR, Sorci G, Riuzzi F, Hsu K, Weber DJ, et al. Functions of S100 Proteins. Curr Mol Med. 2013;13(1):24–57. Sun L, Su Y, Jiao A, Wang X, Zhang B. T cells in health and disease. Signal Transduct Target Ther. 2023 June;19(1):235. Rodriguez OL, Safonova Y, Silver CA, Shields K, Gibson WS, Kos JT et al. Genetic variation in the immunoglobulin heavy chain locus shapes the human antibody repertoire. Nat Commun. 2023 July 21;14(1):4419. Elenbaas JS, Jung IH, Coler-Reilly A, Lee PC, Alisio A, Stitziel NO. The emerging Janus face of SVEP1 in development and disease. Trends Mol Med. 2023;29(11):939–50. Li H, Huang C, Su R, Wang M, Ma Y, Wang Y, et al. Developing a Panel of Shared Susceptibility Genes as Diagnostic Biomarkers for chronic obstructive pulmonary disease and Heart Failure. Comput Biol Med. 2025 Sept;196:110657. Lotfinaghsh A, Imam A, Pompian A, Stitziel NO, Javaheri A. Clinical Insights from Proteomics in Heart Failure. Curr Heart Fail Rep. 2025;22(1):12. Hu C, Yang J, Qi Z, Wu H, Wang B, Zou F, et al. Heat shock proteins: Biological functions, pathological roles, and therapeutic opportunities. MedComm. 2022;3(3):e161. Zuehlke AD, Beebe K, Neckers L, Prince T. Regulation and function of the human HSP90AA1 gene. Gene. 2015;570(1):8–16. Xie Y, Li HF, Sun L, Kusner LL, Wang S, Meng Y, et al. The Role of Osteopontin and Its Gene on Glucocorticoid Response in Myasthenia Gravis. Front Neurol. 2017;8:230. Kusner LL, Kaminski HJ, Soltys J. Effect of complement and its regulation on myasthenia gravis pathogenesis. Expert Rev Clin Immunol. 2008;4(1):43–52. Kwon YC, Kim H, Meyer K, Di Bisceglie AM, Ray R. Distinct CD55 Isoform Synthesis and Inhibition of Complement-Dependent Cytolysis by Hepatitis C Virus. J Immunol Baltim Md 1950. 2016;197(4):1127–36. Sio YY, Anantharaman R, Lee SQE, Matta SA, Ng YT, Chew FT. The Asthma-associated PER1-like domain-containing protein 1 (PERLD1) Haplotype Influences Soluble Glycosylphosphatidylinositol Anchor Protein (sGPI-AP) Levels in Serum and Immune Cell Proliferation. Sci Rep. 2020;10(1):715. Weis CA, Aban IB, Cutter G, Kaminski HJ, Scharff C, Grießmann BW, et al. Histopathology of thymectomy specimens from the MGTX-trial: Entropy analysis as strategy to quantify spatial heterogeneity of lymphoid follicle and fat distribution. PLoS ONE. 2018;13(6):e0197435. Kaminski HJ, Denk J. Corticosteroid Treatment-Resistance in Myasthenia Gravis. Front Neurol. 2022;13:886625. Supplementary Files AdditionalFile1.pdf Additional File 1 .pdf Supplemental Figure 1. Greater differences exist between MG patients and controls than can be explained by prednisone (PA) use at the time of sample collection. A. Venn diagram indicating the overlap of proteins that were significantly differently expressed between MG patients on PA vs HC, MG patients not on PA vs HC, and MG patients on vs not on PA. B. Volcano plot of DEPs between MG patients on PA at study enrollment versus MG patients that were not on PA. CCDC80 and LAG3 were downregulated in patients on PA compared to patients that were not, and 21 DEPs were significantly upregulated. C. Volcano plot of DEPs between MG patients not on PA at study enrollment versus healthy controls. 23 DEPs were downregulated and 299 were upregulated in MG patients not on PA compared to HC. D. Volcano plot of DEPs between MG patients on PA at study enrollment healthy controls. 28 DEPs were downregulated and 392 were upregulated in MG compared to HC. AdditionalFile2.pdf Additional File 2 .pdf Supplemental Figure 2. Features selected as important by Boruta were specific to treatment group across iterations. A. PCA of ETTX patients using the 38 proteins identified by the first round Boruta feature selection for PA patients. Each dot represents one patient and is colored according to whether or not they improved. B. PCA of PA patients using the 45 proteins identified by the first round of Boruta feature selection for ETTX patients. C. PCA of ETTX patients using the 7 proteins identified by the second round of Boruta feature selection for PA patients. D. PCA of PA patients using 17 proteins identified by the second round of Boruta feature selection for ETTX patients. AdditionalFile3.pdf Additional File 3 .pdf Supplemental Figure 3. Feature selection for PA patients. A. PCA of 38 proteins identified as important by Boruta in PA patients, colored according to the patients’ prednisone use at the time of enrollment. B. PCA of 38 proteins in PA patients, demonstrating separation of patients by improvement status. C. Representative feature importance plot from 100 runs of Boruta using 38 previously identified proteins out of 1,210. D. A heatmap of Pearson’s correlation (r values) of the top 7 selected proteins, which indicated no correlation patterns between the selected proteins. E. Averaged ROC curves across the nCV outer folds for each model separately. Shading behind the curve indicates that model’s standard deviation for the average ROC curve. The ROC curves are in the order of highest to lowest AUC. F. Rather than averaging the results from each test set of individuals over the five folds, the predictions were pooled together, as each individual was only tested once. The pooled ROC curves are colored by model and their respective AUC are recorded in the legend. G. Pooled ROC curves for each model, presented separately in the order of highest to lowest AUC. AdditionalFile4.pdf Additional File 4 .pdf Supplemental Figure 4. Feature selection for ETTX patients. A. PCA of 45 proteins identified as important by Boruta in ETTX patients, colored according to the patients’ prednisone use at the time of enrollment. B. PCA of 45 proteins in ETTX patients, demonstrating separation of patients by improvement status. C. Representative feature importance plot from 100 runs of Boruta using 45 previously identified proteins out of 1,210. D. A heatmap of Pearson’s correlation (r values) of the top 17 selected proteins, which indicated no correlation patterns between the selected proteins. E. Averaged ROC curves across the nCV outer folds for each model separately. Shading behind the curve indicates that model’s standard deviation for the average ROC curve. The ROC curves are in the order of highest to lowest AUC. F. Rather than averaging the results from each test set of individuals over the five folds, the predictions were pooled together, as each individual was only tested once. The pooled ROC curves are colored by model and their respective AUC are recorded in the legend. G. Pooled ROC curves for each model, presented separately in the order of highest to lowest AUC. AdditionalFile5.xlsx Additional File 5 .xls Supplemental Tables 1-8. Each table is present on its own labelled excel sheet tab. S Table 1. Patient demographics of myasthenia gravis patients and healthy controls at baseline. Values reported as percentage (number of individuals) for categorical factors and average ± SEM (range) for continuous variables. S Table 2. The parameters tested in a grid search for each model, when applicable, during the inner loop of the nested cross validation. S Table 3. R packages and versions used in analyses and figure generation. S Table 4. ETTX and PA patient characteristics at MGTX trial enrollment. Values reported as percentage (number of individuals) for categorical factors and average ± SEM for continuous variables. Wilcoxon rank-sum and Chi-square tests indicated no differences in the reported metrics between treatment groups. S Table 5. A list of the DEGs from Figure 3 Volcano Plots. For each DEG, the log fold change, average expression of the relative abundance, t statistic, p-value, adjusted p-value, and B statistic are reported. S Table 6. First Boruta Run protein names and number of runs out of 100 that the protein was confirmed as important. S Table 7. Complete t-test results of PA comparison of predictive proteins between subjects that improved versus those that did not. Top: PA predictive proteins. Bottom: ETTX predictive proteins as measured in PA subjects. S Table 8. Complete t-test results of ETTX comparison of predictive proteins between subjects that improved versus those that did not. Top: ETTX predictive proteins. Bottom: PA predictive proteins as measured in ETTX subjects. AdditionalFile6.pdf Additional File 6 .pdf GeneAgent analysis of PA predictive proteins. AdditionalFile7.pdf Additional File 7 .pdf GeneAgent analysis of ETTX predictive proteins. Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 07 Jan, 2026 Reviewers invited by journal 14 Nov, 2025 Editor assigned by journal 24 Oct, 2025 First submitted to journal 21 Oct, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7915342","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":544996692,"identity":"3bc13cac-d071-417c-96a4-fc7ac3e31123","order_by":0,"name":"Karli Faith Gilbert","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA60lEQVRIiWNgGAWjYBACCQYGxgMJDAw8fBIgbgVcFK8WBrAWNjDrDLFaQAywloNtRGiRbD/74MADhloZNunmY58/zqtN3HCA+eBtHjxapHnSDYAOO87DJnMsecbBbceBWtiSrfFpkWNIA/nlGNAvOcYMB7cdy91wgMdMGq8W/mcwLfmfGQ7OAWnh/4ZXi7QE2JYakC3MDAcbakC2sOHVIjkDZIvBAZBfjBnOHDtQP/Mwm7HlHDxaJM6nMT78UVFnzy/d/JihoqbOmO9488Mbb/BogQCDwzAWkMFMUDkY1GEwRsEoGAWjYBTAAQBSME1pIaTmRgAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0002-3055-5213","institution":"George Washington University","correspondingAuthor":true,"prefix":"","firstName":"Karli","middleName":"Faith","lastName":"Gilbert","suffix":""},{"id":544996693,"identity":"b8af4eba-92a5-4847-9bcc-634a61783516","order_by":1,"name":"Amrita K. Cheema","email":"","orcid":"","institution":"Georgetown University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Amrita","middleName":"K.","lastName":"Cheema","suffix":""},{"id":544996694,"identity":"ee37d0da-d01d-44bd-b066-163ebbd3a587","order_by":2,"name":"Henry Kaminski","email":"","orcid":"","institution":"George Washington University","correspondingAuthor":false,"prefix":"","firstName":"Henry","middleName":"","lastName":"Kaminski","suffix":""},{"id":544996695,"identity":"c4940027-a006-439f-aabb-dd4426977292","order_by":3,"name":"Linda Kusner","email":"","orcid":"https://orcid.org/0000-0002-5530-2140","institution":"George Washington University","correspondingAuthor":false,"prefix":"","firstName":"Linda","middleName":"","lastName":"Kusner","suffix":""}],"badges":[],"createdAt":"2025-10-21 16:40:08","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7915342/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7915342/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":96916572,"identity":"37852869-41fd-4141-9c58-5b32dbc8c3d3","added_by":"auto","created_at":"2025-11-27 14:08:44","extension":"xml","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11801,"visible":true,"origin":"","legend":"","description":"","filename":"jtrmJTRMD2518501.xml","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/eebf611b5ad3a4a6e56c508c.xml"},{"id":96792855,"identity":"f192d5b1-66ca-409e-b065-bf5b543a1433","added_by":"auto","created_at":"2025-11-26 07:06:07","extension":"xml","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1439,"visible":true,"origin":"","legend":"","description":"","filename":"JTRMD2518501151489.go.xml","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/7b69bf7dd158b86e785e9753.xml"},{"id":96792860,"identity":"8d243d42-c933-4065-ab88-7dbb9e5184d3","added_by":"auto","created_at":"2025-11-26 07:06:07","extension":"xml","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":823,"visible":true,"origin":"","legend":"","description":"","filename":"JTRMD2518501Import.xml","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/c9f963d50f89df1e35d81b29.xml"},{"id":96917458,"identity":"5d86c2de-27e5-4c9d-a9aa-5e34dde62942","added_by":"auto","created_at":"2025-11-27 14:09:45","extension":"xml","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":202513,"visible":true,"origin":"","legend":"","description":"","filename":"JTRMD25185010enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/b10a685c5177efa55ef2cc97.xml"},{"id":96792877,"identity":"e97070a0-fbf2-435f-9046-c6ba82babc6d","added_by":"auto","created_at":"2025-11-26 07:06:07","extension":"pdf","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3228638,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/6604f0d0fe1b36f69a69c2b1.pdf"},{"id":96792868,"identity":"f746a046-0264-4491-ba5c-26a51b8fce21","added_by":"auto","created_at":"2025-11-26 07:06:07","extension":"pdf","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1536737,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/46c3499ee8d4a8d96aa7e6f2.pdf"},{"id":96792869,"identity":"9c75a961-28b9-4967-84b5-5f4ceefe62ff","added_by":"auto","created_at":"2025-11-26 07:06:07","extension":"pdf","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1572117,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/d37d74a2dab9a552813911d3.pdf"},{"id":96916554,"identity":"7ed2f873-336e-434b-adab-13759462621b","added_by":"auto","created_at":"2025-11-27 14:08:42","extension":"pdf","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":888310,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/1c60b2d5c6a0ba5309bd0a11.pdf"},{"id":96792872,"identity":"a500c8ee-e6b7-4713-b2b5-1130307c56f1","added_by":"auto","created_at":"2025-11-26 07:06:07","extension":"pdf","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1066511,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/fb77dd89057f6ec891c6cb11.pdf"},{"id":96792874,"identity":"8047b0ed-e4f9-43d1-9ca2-05cbbe7bb61c","added_by":"auto","created_at":"2025-11-26 07:06:07","extension":"pdf","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":729839,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/4b4f5a23b478d90c33e0c89f.pdf"},{"id":96916694,"identity":"c3d129f3-9706-4fe4-897f-08b867dafe92","added_by":"auto","created_at":"2025-11-27 14:08:50","extension":"xml","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":199161,"visible":true,"origin":"","legend":"","description":"","filename":"JTRMD25185010structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/ca0228485895d4b0e6dd6919.xml"},{"id":96792876,"identity":"612d9b49-37a8-4f60-8846-e87dfece18a3","added_by":"auto","created_at":"2025-11-26 07:06:07","extension":"html","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":215176,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/d5e49a6cd421baa9cd5099ec.html"},{"id":96792852,"identity":"9590f5ce-948a-4b26-ba52-7457e87ed9b4","added_by":"auto","created_at":"2025-11-26 07:06:07","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":324553,"visible":true,"origin":"","legend":"\u003cp\u003ePatients with MG have distinct serum proteomic profiles compared to non-autoimmune controls. A. Volcano plot of 1,210 detected serum proteins in MG patients relative to controls. Significant proteins (red) are those that passed cutoffs (dotted lines) of \u0026gt; 1 or \u0026lt; -1 log2 fold change and less than the -log10 adjusted p-value of 0.05. B. PCA plot of all patients using all 1,210 proteins. Ellipses represent corresponding 95% confidence intervals for each healthy controls (pink), MG patients on prednisone (blue), and MG patients not on prednisone (dark blue) at the time of study enrollment. C. Unsupervised clustered heatmap of 398 DEPs (Euclidean distance and complete-linkage clustering). Heatmap was created using \u003cem\u003epheatmap\u003c/em\u003e.\u003c/p\u003e","description":"","filename":"fig1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/9822a7a72c10641f065bd8c7.jpg"},{"id":96916494,"identity":"5b200eaf-2db1-4ecb-8be8-ca7c4641273a","added_by":"auto","created_at":"2025-11-27 14:08:39","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":625390,"visible":true,"origin":"","legend":"\u003cp\u003ePathway enrichment analyses indicate significant changes in immune relevant pathways. The order of the dots is presented according to the pathway gene ratio score, the color represents the Benjamini-Hochberg (BH) adjusted p-value, and the size represents the number of genes from the dataset that matched the pathway annotation. A. Dot plot of the 25 most significant Over Representation Analysis (ORA) of the Gene Ontology Biological Process (GO: BP) pathways. B. Dot plot of the 25 most significant ORA of Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways. C. GSEA dot plot of the top 25 most significantly enriched GO-BP pathways in MG patients. D. GSEA score plots immunoglobulin production (GO:0002377, p = 7.9e-5), complement activation (GO:0006956, p = 0.009), and T cell receptor signaling pathway (GO:0050852, p = 0.03).\u003c/p\u003e","description":"","filename":"fig2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/d29a24a2da2be9225e534058.jpg"},{"id":96917250,"identity":"05084320-bcfd-4339-b34d-170d3d23de37","added_by":"auto","created_at":"2025-11-27 14:09:26","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":299420,"visible":true,"origin":"","legend":"\u003cp\u003eBaseline ETTX and PA differences between those who improved versus those who did not. Significance threshold of p-value \u0026lt; 0.5 and |log-fold change| \u0026gt; 1. A. PCA plot of all MG patients using all 1,210 proteins. Colors indicate if the patient improved in their QMG score at 6 months or not. B. Volcano plot of all proteins in patients that improve at 6 months in relation to those who did not improve. C. PCA plot of all MG patients, using all proteins, colored according to treatment group assignment, prednisone alone (blue) or thymectomy plus prednisone (pink). D. Volcano plot of all proteins in patients assigned to thymectomy plus prednisone versus patients receiving prednisone alone. E. Volcano plot of all proteins from ETTX patients who improved compared to ETTX patients that did not improve. F. Volcano plot of all proteins from PA patients who improved compared to PA patients that did not improve.\u003c/p\u003e","description":"","filename":"fig3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/a9440e8509de781fdeabcfac.jpg"},{"id":96917788,"identity":"235b12c4-41c9-4e42-a60c-07007ac8533c","added_by":"auto","created_at":"2025-11-27 14:10:34","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":205785,"visible":true,"origin":"","legend":"\u003cp\u003eFeature selection and nested cross validation workflow.\u003c/p\u003e","description":"","filename":"fig4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/1bcbcb048252dccfa9d26e22.jpg"},{"id":96792853,"identity":"87e8c298-0067-436d-8bc9-d8d37d3bd432","added_by":"auto","created_at":"2025-11-26 07:06:07","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":280236,"visible":true,"origin":"","legend":"\u003cp\u003eRandom Forest model of 7 proteins in PA-treated patients was the best predictor of improvement. A. Comparison of ROC curves between 8 different machine learning models, averaged across five outer-folds. Model’s AUC mean and SEM are reported in the legend. The dotted gray diagonal line represents hypothetical performance of a model that classifies at random. B. Heatmap of p-values from DeLong’s test, which compares if the AUCs between two ROC curves generated from different ML model predictions are statistically different from one another. C. Confusion matrix of the random forest algorithm, using summed classification results across the outer five-folds from the nCV. D. Feature importance scores histogram of the 7 predictive proteins (mean ± SEM). Each dot represents the importance score from one individual outer fold. CCL16 was not deemed important in 2 of the outer folds and RPS3 in 1 of the outer folds, which is why they each have 3 and 4 dots on the graph, respectively.\u003c/p\u003e","description":"","filename":"fig5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/c49ae2eeb19e50afc990b0e7.jpg"},{"id":96916188,"identity":"2090aebe-de2e-410a-a95a-7b064e4f1496","added_by":"auto","created_at":"2025-11-27 14:08:10","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":368362,"visible":true,"origin":"","legend":"\u003cp\u003eExtreme Gradient Boosting model of 17 proteins was best predicting improvement in ETTX-treated patients. A. Comparison of ROC curves between 8 different machine learning models, averaged across five outer-folds. Model’s AUC mean and SEM are reported in the legend. The dotted gray diagonal line represents hypothetical performance of a model that classifies at random. B. Heatmap of p-values from DeLong’s test, which compares if the AUCs between two ROC curves are statistically different from one another. C. Confusion matrix of the XGBoost algorithm, using summed results across the outer five-folds from the nCV. D. SHAP bee swarm plot for the XGBoost model, providing a visual explanation of how the most influential proteins contributed to the model’s predictions. Each dot is representative of one patient and its color corresponds to the value of that protein. E. SHAP importance plot of the features ranked from most to least important. The mean absolute value of the SHAP value is reported.\u003c/p\u003e","description":"","filename":"fig6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/24606fc4f7bdaa9444fc498f.jpg"},{"id":97136377,"identity":"20254e17-e0cd-4fb0-a2b4-34aca4fb368d","added_by":"auto","created_at":"2025-12-01 09:56:29","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3328711,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/8f645b19-c208-49b7-abb8-bf9ba2079331.pdf"},{"id":96916759,"identity":"9954a149-d880-4ed9-9311-e6e44dfcf3cb","added_by":"auto","created_at":"2025-11-27 14:08:52","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1441561,"visible":true,"origin":"","legend":"\u003cp\u003eAdditional File 1 .pdf\u003c/p\u003e\n\u003cp\u003eSupplemental Figure 1. Greater differences exist between MG patients and controls than can be explained by prednisone (PA) use at the time of sample collection. A. Venn diagram indicating the overlap of proteins that were significantly differently expressed between MG patients on PA vs HC, MG patients not on PA vs HC, and MG patients on vs not on PA. B. Volcano plot of DEPs between MG patients on PA at study enrollment versus MG patients that were not on PA. CCDC80 and LAG3 were downregulated in patients on PA compared to patients that were not, and 21 DEPs were significantly upregulated. C. Volcano plot of DEPs between MG patients not on PA at study enrollment versus healthy controls. 23 DEPs were downregulated and 299 were upregulated in MG patients not on PA compared to HC. D. Volcano plot of DEPs between MG patients on PA at study enrollment healthy controls. 28 DEPs were downregulated and 392 were upregulated in MG compared to HC.\u003c/p\u003e","description":"","filename":"AdditionalFile1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/65367fcf6d6f8570e94c40e3.pdf"},{"id":96792858,"identity":"7da4ba4b-cae0-43f1-afa5-b3a44b54f4b5","added_by":"auto","created_at":"2025-11-26 07:06:07","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":494845,"visible":true,"origin":"","legend":"\u003cp\u003eAdditional File 2 .pdf\u003c/p\u003e\n\u003cp\u003eSupplemental Figure 2. Features selected as important by Boruta were specific to treatment group across iterations. A. PCA of ETTX patients using the 38 proteins identified by the first round Boruta feature selection for PA patients. Each dot represents one patient and is colored according to whether or not they improved. B. PCA of PA patients using the 45 proteins identified by the first round of Boruta feature selection for ETTX patients. C. PCA of ETTX patients using the 7 proteins identified by the second round of Boruta feature selection for PA patients. D. PCA of PA patients using 17 proteins identified by the second round of Boruta feature selection for ETTX patients.\u003c/p\u003e","description":"","filename":"AdditionalFile2.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/0a82ff612ad6964a5a1f6846.pdf"},{"id":96792864,"identity":"d4ab250b-6053-47ae-9aff-31d4835d49d6","added_by":"auto","created_at":"2025-11-26 07:06:07","extension":"pdf","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":1196968,"visible":true,"origin":"","legend":"\u003cp\u003eAdditional File 3 .pdf\u003c/p\u003e\n\u003cp\u003eSupplemental Figure 3. Feature selection for PA patients. A. PCA of 38 proteins identified as important by Boruta in PA patients, colored according to the patients’ prednisone use at the time of enrollment. B. PCA of 38 proteins in PA patients, demonstrating separation of patients by improvement status. C. Representative feature importance plot from 100 runs of Boruta using 38 previously identified proteins out of 1,210. D. A heatmap of Pearson’s correlation (r values) of the top 7 selected proteins, which indicated no correlation patterns between the selected proteins. E. Averaged ROC curves across the nCV outer folds for each model separately. Shading behind the curve indicates that model’s standard deviation for the average ROC curve. The ROC curves are in the order of highest to lowest AUC. F. Rather than averaging the results from each test set of individuals over the five folds, the predictions were pooled together, as each individual was only tested once. The pooled ROC curves are colored by model and their respective AUC are recorded in the legend. G. Pooled ROC curves for each model, presented separately in the order of highest to lowest AUC.\u003c/p\u003e","description":"","filename":"AdditionalFile3.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/143b6f4b0c866bba4c583f07.pdf"},{"id":96916810,"identity":"bc5d9ef0-40b5-43b7-9b3d-cadc3cfd732d","added_by":"auto","created_at":"2025-11-27 14:08:55","extension":"pdf","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":768925,"visible":true,"origin":"","legend":"\u003cp\u003eAdditional File 4 .pdf\u003c/p\u003e\n\u003cp\u003eSupplemental Figure 4. Feature selection for ETTX patients. \u0026nbsp;A. PCA of 45 proteins identified as important by Boruta in ETTX patients, colored according to the patients’ prednisone use at the time of enrollment. B. PCA of 45 proteins in ETTX patients, demonstrating separation of patients by improvement status. C. Representative feature importance plot from 100 runs of Boruta using 45 previously identified proteins out of 1,210. D. A heatmap of Pearson’s correlation (r values) of the top 17 selected proteins, which indicated no correlation patterns between the selected proteins. E. Averaged ROC curves across the nCV outer folds for each model separately. Shading behind the curve indicates that model’s standard deviation for the average ROC curve. The ROC curves are in the order of highest to lowest AUC. F. Rather than averaging the results from each test set of individuals over the five folds, the predictions were pooled together, as each individual was only tested once. The pooled ROC curves are colored by model and their respective AUC are recorded in the legend. G. Pooled ROC curves for each model, presented separately in the order of highest to lowest AUC.\u003c/p\u003e","description":"","filename":"AdditionalFile4.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/da0720c692a4f046fe29e6a7.pdf"},{"id":96916497,"identity":"4c9b9022-169d-467f-8225-478e5e1c2e59","added_by":"auto","created_at":"2025-11-27 14:08:39","extension":"xlsx","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":52410,"visible":true,"origin":"","legend":"\u003cp\u003eAdditional File 5 .xls\u003c/p\u003e\n\u003cp\u003eSupplemental Tables 1-8. Each table is present on its own labelled excel sheet tab.\u003c/p\u003e\n\u003cp\u003eS Table 1. Patient demographics of myasthenia gravis patients and healthy controls at baseline. Values reported as percentage (number of individuals) for categorical factors and average ± SEM (range) for continuous variables.\u003c/p\u003e\n\u003cp\u003eS Table 2. The parameters tested in a grid search for each model, when applicable, during the inner loop of the nested cross validation.\u003c/p\u003e\n\u003cp\u003eS Table 3. R packages and versions used in analyses and figure generation.\u003c/p\u003e\n\u003cp\u003eS Table 4. ETTX and PA patient characteristics at MGTX trial enrollment. Values reported as percentage (number of individuals) for categorical factors and average ± SEM for continuous variables. Wilcoxon rank-sum and Chi-square tests indicated no differences in the reported metrics between treatment groups.\u003c/p\u003e\n\u003cp\u003eS Table 5. A list of the DEGs from Figure 3 Volcano Plots. For each DEG, the log fold change, average expression of the relative abundance, t statistic, p-value, adjusted p-value, and B statistic are reported.\u003c/p\u003e\n\u003cp\u003eS Table 6. First Boruta Run protein names and number of runs out of 100 that the protein was confirmed as important.\u003c/p\u003e\n\u003cp\u003eS Table 7. Complete t-test results of PA comparison of predictive proteins between subjects that improved versus those that did not. Top: PA predictive proteins. Bottom: ETTX predictive proteins as measured in PA subjects.\u003c/p\u003e\n\u003cp\u003eS Table 8. Complete t-test results of ETTX comparison of predictive proteins between subjects that improved versus those that did not. Top: ETTX predictive proteins. Bottom: PA predictive proteins as measured in ETTX subjects.\u003c/p\u003e","description":"","filename":"AdditionalFile5.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/35cd54fe665c11f9b258c891.xlsx"},{"id":96792865,"identity":"d6774629-5d95-4d3d-9be1-d0f19e1a0bc3","added_by":"auto","created_at":"2025-11-26 07:06:07","extension":"pdf","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":148197,"visible":true,"origin":"","legend":"\u003cp\u003eAdditional File 6 .pdf\u003c/p\u003e\n\u003cp\u003eGeneAgent analysis of PA predictive proteins.\u003c/p\u003e","description":"","filename":"AdditionalFile6.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/9fbd7238915faffb3849d9fc.pdf"},{"id":96916800,"identity":"eb0e2788-582e-423f-b8c0-77aed61be53e","added_by":"auto","created_at":"2025-11-27 14:08:53","extension":"pdf","order_by":7,"title":"","display":"","copyAsset":false,"role":"supplement","size":190662,"visible":true,"origin":"","legend":"\u003cp\u003eAdditional File 7 .pdf\u003c/p\u003e\n\u003cp\u003eGeneAgent analysis of ETTX predictive proteins.\u003c/p\u003e","description":"","filename":"AdditionalFile7.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7915342/v1/d50a2c4a2b32d32245d2c451.pdf"}],"financialInterests":"","formattedTitle":"Proteomic and Machine Learning Analysis Predicts Treatment Response Signatures in Myasthenia Gravis","fulltext":[{"header":"Background","content":"\u003cp\u003eTherapeutic decision-making for patients with autoimmune diseases is often hindered by the central challenge of the unpredictability of treatment response. Whether due to intrinsic disease heterogeneity or variable sensitivity to immunomodulatory therapies, patients with the same diagnosis can experience markedly different outcomes. The complex interplay of many factors, including sex, age, genetic background, and environmental experiences, collectively influence disease expression and therapeutic efficacy. Myasthenia gravis (MG), a prototypical antibody-mediated disorder, exemplifies this clinical variability and offers a compelling model in which to investigate the biological underpinnings of treatment response and identify biomarkers to predict therapeutic response (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eDespite an array of groundbreaking research in the field of MG, few studies have combined rigorous, long-term clinical outcome monitoring with systematic collection of linked biological specimens. A notable exception is the NIH-supported, Thymectomy Trial in Non-Thymomatous Myasthenia Gravis Patients Receiving Prednisone Therapy (MGTX) trial, which incorporated comprehensive, validated clinical assessments linked with prospectively collected biospecimens. The MGTX trial, with a primary endpoint at 3 years, established the efficacy of thymectomy plus prednisone (ETTX) compared with prednisone alone (PA) in patients aged 18 to 65 years with acetylcholine receptor (AChR) antibody-positive MG (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). Subjects in the thymectomy arm had a lower total Quantitative MG (QMG) score and required lower overall corticosteroid dosage at end of study, although no significant differences were detected at 6 months. Prednisone is a well-established therapy for MG (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e), while the application of thymectomy dates back nearly 100 years, but was only unequivocally identified as effective by the MGTX trial. However, the mechanisms underlying the benefit of thymectomy remain poorly understood. Considerable variability in treatment response was observed across both study arms.\u003c/p\u003e\u003cp\u003eBroad-based -omics profiling approaches provide a powerful method to move beyond clinical observation toward molecular signatures that predict therapeutic outcomes. Optimization of treatment regimens in patients with MG can be supported by the identification of biomarkers predictive of treatment outcomes (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e), a task increasingly enabled by machine learning approaches applied to clinical features and multi-omics data. Studies applying machine learning to investigate biospecimens from MG patients are limited and have used publicly available microarray expression profiles (\u003cspan additionalcitationids=\"CR9\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e), while others have generated primary proteomic (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e), metabolomic (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e), and transcriptomic data (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e) from MG cohorts.\u003c/p\u003e\u003cp\u003eHere, we applied liquid chromatography\u0026ndash;mass spectrometry (LC-MS) to sera collected at study entry (baseline) from 86 subjects of MGTX trial, of whom 41 were assigned to PA and 45 to ETTX. We first assessed for differences between MG subjects and controls and then aimed to identify serum proteins at baseline predictive of 6-month QMG improvement in treatment groups. Although subjects who had undergone thymectomy demonstrated a greater average reduction in QMG scores compared to prednisone-only subjects, about 30 percent of subjects did not achieve a well-established criterion of a clinically meaningful reduction of 3 points on the QMG score in the first 6 months of the study. We chose this early time point because of the greater uniformity of treatment early in the study based on uniform prednisone treatment protocol, an absence of rescue therapies, and expectation of thymectomy providing an early effect on the autoimmune pathology of MG. We made no prior assumptions regarding the data set and applied several machine learning (ML) models for predictors of treatment response.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eStudy Population\u003c/h2\u003e\u003cp\u003eSerum samples were obtained from participants enrolled in the MGTX clinical trial (NS42685). Specimens had been stored at -80\u003csup\u003eo\u003c/sup\u003eC at NCI-Fredrick and then transferred to the biospecimen bank at George Washington University. Samples were collected at the baseline time point for each participant and analyzed using LC-MS. Clinical outcome measurements were collected at baseline and six months using the QMG assessment. All individuals included in this analysis had confirmed diagnosis of MG with detectable acetylcholine-receptor autoantibodies and no evidence of thymoma (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). Control serum samples were collected at George Washington University with the exclusion criteria of autoimmune disease and vaccination in the past month. All participants were older than 18 years and provided written informed consent. Patient and control characteristics are summarized in Supplemental Table\u0026nbsp;1.\u003c/p\u003e\u003cp\u003eThe NINDS funded the trial and assembled an independent Data Safety Monitoring Board. Sites received local institutional review board (IRB)/ethics committee approvals, and each patient provided written informed consent before study entry including provision of serum samples. All specimens were deidentified. The George Washington University IRB provided additional review and approved use of specimens.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eMass Spectrometry Protein Analysis\u003c/h3\u003e\n\u003cp\u003eShotgun proteomics of MG patient sera was performed as previously described (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e). Individual samples (20 \u0026micro;L) and a pooled quality control (5 \u0026micro;L of each sample) were prepared using ENRICH-iST Kit (PreOmics\u0026reg;) before injection into the mass spectrometer nanoElute\u0026reg; 2 instrument (Bruker Scientific). Peptide fractions from the pooled quality control (9 technical replicates) were obtained using the Pierce\u0026trade; High pH Reversed-Phase Peptide Fractionation Kit. A data dependent acquisition (DDA) spectral library from the peptide fractions was generated in Fragpipe and DIA_SpectLib_quant workflow. Individual samples were quantified with DIA-NN (data independent acquisition \u0026ndash; neural network) with a maximum false discovery rate of 1%.\u003c/p\u003e\n\u003ch3\u003eData Cleaning and Imputation\u003c/h3\u003e\n\u003cp\u003eThe DIA quantification values were first filtered in Fragpipe (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e) to remove contaminant proteins, reverse sequences, and proteins identified by a single peptide. The data were then converted to a log2 scale and separated by group (MG vs Control). Missing values were imputed using a Perseus-type Missing Not at Random (MNAR) approach, where values were randomly drawn from a left-shifted Gaussian distribution. This distribution was centered 1.8 standard deviations below the mean of the observed values within each group, with a standard deviation width of 0.3, to simulate low-abundance proteins below the detection limit.\u003c/p\u003e\n\u003ch3\u003eAnalysis of Myasthenia Gravis versus Control Subjects\u003c/h3\u003e\n\u003cp\u003eThe R package \u003cem\u003elimma\u003c/em\u003e was used to determine the differentially expressed proteins (DEPs) between MG and healthy control (HC) samples (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e). DEPs were defined as proteins that surpassed a significance cutoff of a |log\u003csub\u003e2\u003c/sub\u003e fold change| \u0026gt;1 and an adjusted p-value less than 0.05 (Benjamini-Hochberg method). Pathway enrichment analyses were performed using over representation analysis (ORA) and Gene Set Enrichment Analysis (GSEA) with the R packages \u003cem\u003eclusterProfiler\u003c/em\u003e and \u003cem\u003eenrichPlot\u003c/em\u003e (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e).\u003c/p\u003e\n\u003ch3\u003eFeature Selection\u003c/h3\u003e\n\u003cp\u003eWe applied the Boruta feature selection algorithm to identify all relevant variables for predictive detection of treatment predictive proteins. The approach enables optimization of a low sample/feature ratio, to reduce the number of proteins, and thus noise, in our predictive analyses. This allowed for enhanced performance in downstream assessments and reduced the likelihood of a classification algorithm from overfitting a model (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e).\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eBoruta\u003c/h2\u003e\u003cp\u003eThe Boruta algorithm (\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e) was employed in R using the \u003cem\u003eBoruta\u003c/em\u003e package with default settings, \u003cem\u003emaxRuns\u003c/em\u003e\u0026thinsp;=\u0026thinsp;300, and results recorded from 100 iterations, to obtain a subset of proteins predictive of clinical improvement in thymectomy plus prednisone (ETTX) and prednisone alone (PA) subjects separately. In each iteration, Boruta creates shadow features that are added to the dataset, runs a single random forest model on the dataset and compares the importance scores of the real variables to the shadow features. Features whose importance scores were significantly higher than the highest importance score observed among the random noise variables were retained as relevant predictors. The initial subsets of proteins were then subjected to a second set of 100 iterations of Boruta feature selection to obtain a set of proteins for PA and ETTX patients for use in nested cross validation of eight ML models.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eNested Cross Validation\u003c/h3\u003e\n\u003cp\u003eNested cross-validation (nCV) was employed to separate model training from testing, ensuring unbiased performance estimates and reducing the risk of overfitting in clinical outcome predictions. We separately applied a 5-fold nCV to the selected features for ETTX and PA subjects. The inner loop performed hyperparameter tuning via grid search (Supplemental Table\u0026nbsp;2) using the \u003cem\u003ecaret\u003c/em\u003e R package with its default 10-fold CV to determine the most accurate training parameters for the outer loop. The data were split across the outer folds such that each patient was tested once. Classification predictions and performance metrics were pooled or averaged across outer folds to generate unbiased estimate.\u003c/p\u003e\n\u003ch3\u003enCV Models\u003c/h3\u003e\n\u003cp\u003eTo develop robust and generalizable predictive models, we employed a diverse panel of supervised ML algorithms within the nCV framework: probabilistic classifiers (Na\u0026iuml;ve Bayes, Logistic Regression), regularized linear models (LASSO, Elastic Net), tree-based approaches (Classification Tree, Random Forest), and ensemble boosting methods (AdaBoost, XGBoost). Each model captured different aspects of data structure, which allowed us to comprehensively evaluate predictive features across varying assumptions and model complexities. This broad approach was chosen to minimize bias from any single modeling strategy and to better characterize proteomic expression patterns relevant to predicting clinical improvement.\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eProbabilistic Classifiers\u003c/h2\u003e\u003cp\u003eWe implemented Na\u0026iuml;ve Bayes and logistic regression as representative probabilistic classifiers. Na\u0026iuml;ve Bayes, which assumes independence among features, was applied using the \u003cem\u003enaiveBayes\u003c/em\u003e function from the \u003cem\u003ee1071\u003c/em\u003e R package (\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e). Logistic regression, which models the log-odds of the binary outcome as a linear combination of predictors, was performed using the base R \u003cem\u003eglm\u003c/em\u003e function with a binomial family. Both models are deterministic and reproducible, in contrast to stochastic tree-based and ensemble methods.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003eRegularized Linear Models\u003c/h2\u003e\u003cp\u003eLASSO (Least Absolute Shrinkage and Selection Operator) is a linear regression technique that uses L1 regularization (\u003cem\u003eα\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1) to shrink the coefficients of less important features to zero. Similarly, elastic net is a regularized regression that balances L1 and L2 regularization (0\u0026thinsp;\u0026lt;\u0026thinsp;\u003cem\u003eα\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;1). We performed LASSO and Elastic Net using the \u003cem\u003eglmnet\u003c/em\u003e R package (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e). Using the built-in cross-validation, proteins with non-zero coefficients were recorded from two regularization levels: \u003cem\u003eλ\u003c/em\u003emin, the value of \u003cem\u003eλ\u003c/em\u003e that minimized the cross-validation error, and \u003cem\u003eλ\u003c/em\u003e1se, the largest λ within one standard error of \u003cem\u003eλ\u003c/em\u003emin. These correspond to models that favor predictive accuracy and model simplicity, respectively.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003eDecision Trees\u003c/h2\u003e\u003cp\u003eDeterministic decision trees were made using the classification and regression tree algorithms (CART) from the recursive partitioning and regression tree R package \u003cem\u003erpart\u003c/em\u003e (\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e). For each split at a node, the relative impurity reduction was calculated as the impurity at the parent node minus the total impurity of the left and right child. A named numeric vector for each tree provided the protein names of those considered important by the algorithm.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003eRandom Forests\u003c/h2\u003e\u003cp\u003eThe random forests algorithm was performed using the R package \u003cem\u003erandomForest\u003c/em\u003e (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e), which ranks features based on how much they contribute to decreasing the variance in the model, also referred to as the increase in node purity. The importance scores calculated for each protein are a normalized calculation of the contribution of each protein to the reduction of impurity in the model.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003eEnsemble Boosting\u003c/h2\u003e\u003cp\u003eAdaptive boosting (AdaBoost) and extreme gradient boosting (XGBoost) were used as ensemble learning methods, which work to improve predictive performance by sequentially training models, each focused on correcting the errors of the previous. AdaBoost was implemented using the \u003cem\u003eada\u003c/em\u003e function from the \u003cem\u003eJOUSBoost\u003c/em\u003e R package (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e), while XGBoost was applied using the \u003cem\u003exgboost\u003c/em\u003e package (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e). Both methods iteratively reweigh training observations to emphasize harder-to-classify cases. A key distinction is that AdaBoost used simple decision stumps as base learners, whereas XGBoost built full decision trees, allowing it to capture more complex patterns.\u003c/p\u003e\u003cp\u003eA complete list of R packages and versions used can be found in Supplementary Table\u0026nbsp;3. Statistical analyses were performed in R v4.4.1 and GraphPad Prism 10.3.0. The repository with all R source code used to run nCV is available on GitHub at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/drkgil/NestCV\u003c/span\u003e\u003cspan address=\"https://github.com/drkgil/NestCV\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003ePathway and Tissue Source Analysis\u003c/h2\u003e\u003cp\u003eWe performed GeneAgent analysis (\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e) to assess biological pathways identified for proteins found to be predictive for outcomes for PA and ETTX groups.\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cb\u003eProteomic profiling reveals distinctions between MG patients and controls.\u003c/b\u003e\u003c/p\u003e\u003cp\u003eEighty-six sera at study entrance to the MGTX clinical trial and 37 healthy controls were used (demographic details in Supplemental Table\u0026nbsp;1). While efforts were made to match the groups, the average age of subjects differed from controls (Control: 43.46 years; MG: 35.91 years; Mann-Whitney U\u0026thinsp;=\u0026thinsp;1090, p\u0026thinsp;=\u0026thinsp;0.0053); there was no difference in gender distribution. We identified proteomic signatures differentiating MG patients from controls. Liquid chromatography mass spectrometry analysis identified 1,213 proteins, from which a dataset of 1,210 were used to identify 398 differentially expressed proteins (DEPs). Among these, 23 proteins were significantly downregulated in MG compared to controls, while 375 proteins were significantly upregulated, highlighting a substantial alteration in the proteomic landscape associated with MG (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003ePrincipal component analysis (PCA) of the full proteomic dataset further revealed a clear separation between MG and control samples, emphasizing the robustness of these differences (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eB). Among the MG patients, 77% were on prednisone at the time of study enrollment at an average dose of 30.7 mg\u0026thinsp;\u0026plusmn;\u0026thinsp;14.1 (Supplemental Table\u0026nbsp;1). We found that the use of prednisone at the time of enrollment was not a driver of the PCA separation from controls. Hierarchical clustering of the 398 DEPs further supported this distinction, with heatmap visualization clearly separating MG patients from controls, without a pattern according to prednisone use (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eC).\u003c/p\u003e\u003cp\u003eTo explore the biological processes associated with the 398 DEPs identified in MG, we performed ORA using the Gene Ontology Biological Processes (GO-BP) pathway database. The analysis revealed numerous significantly enriched pathways, with the top 25 pathways included those related to actin filament organization and organelle organization (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA). ORA using the KEGG database mirrored the GO-BP results, with enrichment of terms related to actin organization and cytoskeletal dynamics (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB). Several of these top KEGG pathways corresponded with those typically enriched in MG patients compared to healthy controls: \u0026ldquo;leukocyte transendothelial migration\u0026rdquo;, \u0026ldquo;B cell receptor signaling pathway\u0026rdquo;, and Fc receptor pathways.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo test for coordinated expression patterns that could be informative of changes between MG subjects and controls that may have escaped the threshold of significance set for DEPs, we also performed GSEA utilizing the GO-BP database. Proteins were ranked according to the B-statistic obtained from the \u003cem\u003elimma\u003c/em\u003e analysis of Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. While many of the top pathways overlapped with those identified by ORA(Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC), a few significant pathways emerged. Notably, enrichment scores for immunoglobulin production (p\u0026thinsp;=\u0026thinsp;7.5e-6) and complement activation (p\u0026thinsp;=\u0026thinsp;0.0086) were significantly downregulated in MG patients compared to controls, while the T cell receptor signaling pathway was significantly upregulated (p\u0026thinsp;=\u0026thinsp;0.03) (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eD).\u003c/p\u003e\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\u003ch2\u003eDistinct proteomic signatures emerge when prednisone and thymectomy treatment groups are analyzed separately\u003c/h2\u003e\u003cp\u003eWe next sought to determine which proteins could be a predictive marker of clinical improvement at 6 months based on a greater than three point improvement of the QMG score, which has been established as a clinically meaningful difference (\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e). The patient characteristics at baseline did not differ between the treatment groups (Supplemental Table\u0026nbsp;4).\u003c/p\u003e\u003cp\u003ePCA of MG subject proteomics did not reveal a distinction between patients that improved compared to those that did not improve (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA); however, four proteins were significantly reduced in patients that showed improvement (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB). Similarly, the same PCA labeled by treatment group indicated a lack of clustering (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC), despite DEPs present between ETTX and PA treated patients (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eD). This finding is not surprising given that the clinical trial did not show statistically significant differences between groups until after 6 months.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eNext, we analyzed baseline serum protein profiles separately for each treatment group. Within the ETTX and PA treatment groups, comparisons between patients who improved and those who did not revealed 16 DEPs in ETTX patients (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eE) and 14 DEPs in PA patients (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eF). Given the greater number of DEPs identified within individual treatment groups compared to the combined MG subjects, we maintained this separation in the data analysis.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e\u003ch2\u003eFeature selection highlights proteins predictive of improvement and distinct between treatment groups\u003c/h2\u003e\u003cp\u003eConsidering the positive response of patients to thymectomy, we hypothesized that proteins predictive of clinical improvement would differ between treatment groups. The extent and nature of these differences was initially unclear as potential group-level differences were masked by the abundance of non-significant proteins, as seen in the PCA plots (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA, C).\u003c/p\u003e\u003cp\u003eTo address this, we independently applied Boruta feature selection within each treatment group (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), identifying preliminary sets of 45 proteins in ETTX patients and 38 proteins in PA patients. A second round of feature selection was then applied to these subsets to refine the list to proteins strongly associated with improvement. This two-step selection process reduced dimensional noise, yielding 17 and 7 key proteins for the ETTX and PA groups, respectively (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). At each step of the feature selection, we verified that the sets of features were specific to its respective treatment group by performing PCA with mismatching the patient group and predictive proteins (Supplemental Fig.\u0026nbsp;2). These group-specific protein sets were used in all subsequent analyses to test their performance across eight different machine learning algorithms in a nested cross validation design.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eBaseline proteins predictive of clinical improvement in PA and ETTX treated MG patients.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMG Group\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eProtein\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eImproved\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eNot Improved\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003ep-value\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"6\" rowspan=\"7\"\u003e\u003cp\u003ePA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCCL16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e7.05\u0026thinsp;\u0026plusmn;\u0026thinsp;0.32\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e8.13\u0026thinsp;\u0026plusmn;\u0026thinsp;0.39\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.04\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCILP2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e12.08\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e11.61\u0026thinsp;\u0026plusmn;\u0026thinsp;0.15\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.017\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eGALNT1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e9.30\u0026thinsp;\u0026plusmn;\u0026thinsp;0.30\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e10.27\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.0046\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eIGHV3-43\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e10.73\u0026thinsp;\u0026plusmn;\u0026thinsp;0.22\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e11.30\u0026thinsp;\u0026plusmn;\u0026thinsp;0.16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.04\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRPS3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e8.42\u0026thinsp;\u0026plusmn;\u0026thinsp;0.35\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e7.05\u0026thinsp;\u0026plusmn;\u0026thinsp;0.47\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.025\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRPS3A\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e8.61\u0026thinsp;\u0026plusmn;\u0026thinsp;0.42\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e7.01\u0026thinsp;\u0026plusmn;\u0026thinsp;0.17\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.0012\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eVAPB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e11.46\u0026thinsp;\u0026plusmn;\u0026thinsp;0.23\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e11.95\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.057\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"16\" rowspan=\"17\"\u003e\u003cp\u003eETTX\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eARHGAP9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e7.34\u0026thinsp;\u0026plusmn;\u0026thinsp;0.34\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e8.35\u0026thinsp;\u0026plusmn;\u0026thinsp;0.46\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.086\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCCN5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e9.47\u0026thinsp;\u0026plusmn;\u0026thinsp;0.42\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e8.33\u0026thinsp;\u0026plusmn;\u0026thinsp;0.78\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.21\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCD55\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e6.34\u0026thinsp;\u0026plusmn;\u0026thinsp;0.22\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e7.73\u0026thinsp;\u0026plusmn;\u0026thinsp;0.49\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.019\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCHAD\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e13.06\u0026thinsp;\u0026plusmn;\u0026thinsp;0.15\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e12.38\u0026thinsp;\u0026plusmn;\u0026thinsp;0.21\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.015\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCHRDL1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e9.19\u0026thinsp;\u0026plusmn;\u0026thinsp;0.31\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e7.96\u0026thinsp;\u0026plusmn;\u0026thinsp;0.51\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.051\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eFBLN5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e10.79\u0026thinsp;\u0026plusmn;\u0026thinsp;0.22\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e9.86\u0026thinsp;\u0026plusmn;\u0026thinsp;0.36\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.038\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eHS1BP3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e7.61\u0026thinsp;\u0026plusmn;\u0026thinsp;0.27\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e8.83\u0026thinsp;\u0026plusmn;\u0026thinsp;0.44\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.027\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eHSP90AA1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e9.19\u0026thinsp;\u0026plusmn;\u0026thinsp;0.42\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e10.70\u0026thinsp;\u0026plusmn;\u0026thinsp;0.32\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.0068\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eIGKV1-5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e19.85\u0026thinsp;\u0026plusmn;\u0026thinsp;0.11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e19.61\u0026thinsp;\u0026plusmn;\u0026thinsp;0.19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.30\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMMP3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e12.87\u0026thinsp;\u0026plusmn;\u0026thinsp;0.21\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e13.76\u0026thinsp;\u0026plusmn;\u0026thinsp;0.31\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.026\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePCMT1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e8.49\u0026thinsp;\u0026plusmn;\u0026thinsp;0.34\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e7.88\u0026thinsp;\u0026plusmn;\u0026thinsp;0.67\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.42\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRAB14\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e10.18\u0026thinsp;\u0026plusmn;\u0026thinsp;0.36\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e8.96\u0026thinsp;\u0026plusmn;\u0026thinsp;0.63\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.11\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRDX\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e10.34\u0026thinsp;\u0026plusmn;\u0026thinsp;0.24\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e10.70\u0026thinsp;\u0026plusmn;\u0026thinsp;0.39\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.45\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRNASET2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e11.03\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e10.81\u0026thinsp;\u0026plusmn;\u0026thinsp;0.16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.21\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSELL\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e14.01\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e13.72\u0026thinsp;\u0026plusmn;\u0026thinsp;0.17\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.14\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSPP2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e15.20\u0026thinsp;\u0026plusmn;\u0026thinsp;0.13\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e15.76\u0026thinsp;\u0026plusmn;\u0026thinsp;0.14\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.0059\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSVEP1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e11.06\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e10.02\u0026thinsp;\u0026plusmn;\u0026thinsp;0.26\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.002\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"5\"\u003eValues reported as log\u003csub\u003e2\u003c/sub\u003e abundances (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SEM). P-values are reported from students two-tailed t-tests between improved and not improved groups.\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec20\" class=\"Section2\"\u003e\u003ch2\u003eRandom Forest identifies seven proteins highly predictive of response to prednisone treatment\u003c/h2\u003e\u003cp\u003eWe applied a five-fold nCV framework utilizing eight machine learning models, which decreases the potential for bias in the selected panels, on the relative abundances of 7 proteins, CCL16, CILP2, GALNT1, IGHV3-43, RPS3, RPS3A, and VAPB, expressed in PA patients at baseline. nCV is particularly well-suited for smaller datasets, as it helps prevent overfitting by ensuring each sample is independently tested from the data used to train the model (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eEach model performed relatively well at classifying PA patients as improved or not, except for the classification tree algorithm. Several models showed similar performance, with receiver operating characteristic (ROC) curves and average area under the ROC curve (AUC) closely overlapping (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eA). However, differences between models were more evident when considering the Matthews correlation coefficient (MCC) (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), a robust metric that accounts for all confusion matrix categories. MCC scores range from \u0026minus;\u0026thinsp;1 to 1, indicative of total disagreement and total agreement between predicted and actual classes, respectively (Chicco and Jurman, 2020). Together, the AUC and MCC scores highlight the random forest algorithm as the best-performing classifier for PA patients in the nested cross-validation.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePerformance metrics across 5 outer folds of PA patients.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"9\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMetric\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNB\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eLogistic\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eLASSO\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eEN\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eCT\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eRF\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eAda\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eXGB\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e\u003cem\u003e0.87\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.82\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.87\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.87\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.64\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.91\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e0.74\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e0.82\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAUCPR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.95\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.94\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.94\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.75\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.95\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePrec\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.79\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.80\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.79\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.71\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.86\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e0.83\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e0.77\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.68\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.68\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.80\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.89\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.72\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e0.77\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e0.84\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAcc\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.76\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.68\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.73\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.78\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.64\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.85\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e0.76\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e0.76\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSpec\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.88\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.70\u0026thinsp;\u0026plusmn;\u0026thinsp;0.18\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.65\u0026thinsp;\u0026plusmn;\u0026thinsp;0.19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.62\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.47\u0026thinsp;\u0026plusmn;\u0026thinsp;0.18\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.73\u0026thinsp;\u0026plusmn;\u0026thinsp;0.11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e0.72\u0026thinsp;\u0026plusmn;\u0026thinsp;0.13\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e0.62\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eF\u003csub\u003e1\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.76\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.71\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.78\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.83\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.71\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.88\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e0.80\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e0.80\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMCC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e\u003cem\u003e0.56\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.42\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.52\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.52\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.17\u0026thinsp;\u0026plusmn;\u0026thinsp;0.20\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.68\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e0.47\u0026thinsp;\u0026plusmn;\u0026thinsp;0.17\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e0.47\u0026thinsp;\u0026plusmn;\u0026thinsp;0.14\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"9\"\u003eMetrics reported as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SEM. The best performing AUC and MCC metrics are in bold and the runner up in italics. NB, Na\u0026iuml;ve Bayes; EN, elastic net; CT, classification tree; RF, random forest; Ada, adaptive boosting; XGB, extreme gradient boosting; AUC, area under the ROC curve; AUCPR, area under the precision-recall curve; Prec, precision; Acc, accuracy; Spec, specificity; F\u003csub\u003e1\u003c/sub\u003e, F-score (harmonic mean of precision and recall); MCC, Matthew\u0026rsquo;s correlation coefficient.\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe random forest algorithm produced the highest AUC 0.91\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06, which was significantly better than the logistic regression (Z = -2.48, CI -0.317 to -0.037, p\u0026thinsp;=\u0026thinsp;0.013), extreme gradient boosting (Z\u0026thinsp;=\u0026thinsp;2.89, CI\u0026thinsp;=\u0026thinsp;0.043 to 0.23, p\u0026thinsp;=\u0026thinsp;0.0039), and trended towards being better than adaptive boosting (Z\u0026thinsp;=\u0026thinsp;1.86, CI = -0.006 to 0.216, p\u0026thinsp;=\u0026thinsp;0.063) (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eB). Across the outer folds of the nCV, the RF model correctly classified 35/41 patients (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eC). The mean decrease in accuracy was used to determine which of the 7 proteins were most important to the RF model in accurately predicting patient improvement outcome. The protein \u003cem\u003eVAPB\u003c/em\u003e had the highest average importance score, followed by \u003cem\u003eRPS3A\u003c/em\u003e, and \u003cem\u003eCILP2\u003c/em\u003e (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eD).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e\u003ch2\u003eSeventeen proteins predict treatment response to thymectomy\u003c/h2\u003e\u003cp\u003eSimilarly, we applied five-fold nCV to the 17 proteins predictive of improvement outcome in ETTX patients: ARHGAP9, CCN5, CD55, CHAD, CHRDL1, FBLN5, HS1BP3, HSP90AA1, IGKV1-5, MMP3, PCMT1, RAB14, RDX, RNASET2, SELL, SPP2, and SVEP1. The ROC curve analysis indicated that the adaboost and XGBoost models performed best, followed closely by LASSO, Elastic Net, Random Forest, and Na\u0026iuml;ve Bayes (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eA, Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). The XGBoost model achieved the highest AUC (0.99\u0026thinsp;\u0026plusmn;\u0026thinsp;0.01), performing significantly better than Logistic Regression (Z = -2.67, CI = -0.37 to -0.06, p\u0026thinsp;=\u0026thinsp;0.008) and Classification Tree algorithms (Z = -5.0, CI = -0.68 to -0.30, p\u0026thinsp;=\u0026thinsp;5.7e-7), and trended towards significance over Na\u0026iuml;ve Bayes (Z = -1.80, CI = -0.34 to 0.15, p\u0026thinsp;=\u0026thinsp;0.072) and Random Forest algorithms (Z = -1.35, CI = -0.23 to 0.04, p\u0026thinsp;=\u0026thinsp;0.18) (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eB). Across all five outer folds of the nCV, XGBoost correctly classified 41/45 ETTX patients (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eC).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePerformance metrics across 5 outer folds of ETTX patients.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"9\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMetric\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNB\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eLogistic\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eLASSO\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eEN\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eCT\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eRF\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eAda\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eXGB\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.83\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.66\u0026thinsp;\u0026plusmn;\u0026thinsp;0.13\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.56\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e\u003cem\u003e0.95\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e\u003cb\u003e0.99\u0026thinsp;\u0026plusmn;\u0026thinsp;0.01\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAUCPR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.89\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.96\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.96\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.78\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.96\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e1.00\u0026thinsp;\u0026plusmn;\u0026thinsp;0.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e1.00\u0026thinsp;\u0026plusmn;\u0026thinsp;0.00\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePrec\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.80\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.84\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.76\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.79\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e0.95\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e0.88\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.88\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.76\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.70\u0026thinsp;\u0026plusmn;\u0026thinsp;0.11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.97\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e1.00\u0026thinsp;\u0026plusmn;\u0026thinsp;0.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e1.00\u0026thinsp;\u0026plusmn;\u0026thinsp;0.00\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAcc\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.76\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.71\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.89\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.89\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.58\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.78\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e0.96\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e0.91\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSpec\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.41\u0026thinsp;\u0026plusmn;\u0026thinsp;0.19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.58\u0026thinsp;\u0026plusmn;\u0026thinsp;0.19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.79\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.79\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.42\u0026thinsp;\u0026plusmn;\u0026thinsp;0.19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.44\u0026thinsp;\u0026plusmn;\u0026thinsp;0.23\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e0.83\u0026thinsp;\u0026plusmn;\u0026thinsp;0.11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e0.75\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eF1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.83\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.78\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.91\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.91\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.69\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.85\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e0.97\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e0.93\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMCC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e\u003cp\u003e0.40\u0026thinsp;\u0026plusmn;\u0026thinsp;0.24\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.37\u0026thinsp;\u0026plusmn;\u0026thinsp;0.19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.73\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.73\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.08\u0026thinsp;\u0026plusmn;\u0026thinsp;0.17\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.66\u0026thinsp;\u0026plusmn;\u0026thinsp;0.20\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.88\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c9\"\u003e\u003cp\u003e\u003cb\u003e0.81\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"9\"\u003eMetrics reported as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SEM. The best AUC and MCC metrics are in bold and the runner ups are in italics.\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eSHAP (SHapley Additive exPlanations) values were calculated to explain the XGBoost model. These values correspond to how much the protein contributes to the prediction in each individual; the greater the distribution of a protein\u0026rsquo;s SHAP values, the greater its influence on the prediction of the model. Since the model performs a binary classification for each patient, positive SHAP values correspond to an increased probability of being classified as \u0026lsquo;improved\u0026rsquo;; conversely, negative values correspond to an increased probability of being classified \u0026lsquo;not improved\u0026rsquo; (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eD). In ETTX patients, the XGBoost model identified SVEP1, SPP2, CHAD, and RAB14 as the top important proteins for predicting clinical improvement (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eE).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec22\" class=\"Section2\"\u003e\u003ch2\u003eGeneAgent Analysis\u003c/h2\u003e\u003cp\u003eWe applied GeneAgent (\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e) analysis to the ETTX and PA proteins identified as predictors of clinical outcomes (Supplementary Files 6 \u0026amp; 7). The summary report for ETTX indicated that the extracellular matrix remodeling was the primary pathway involved in treatment response, with key indicators from cell adhesion, migration and immune response. In contrast, analysis of the seven proteins predictive of PA response revealed some association with ribosomal pathways, but no distinct summary was identified.\u003c/p\u003e\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eTo better understand the immunobiological underpinnings of treatment response in MG, we performed serum proteomic profiling of participants in the MGTX clinical trial and healthy controls. We identified distinct serum proteomic signatures that differentiated individuals with MG from controls, with biological pathways involving complement activation, immunoglobulin production, and T cell signaling driving these differences. Moreover, we discovered a subset of proteins that predicted clinical improvement six months later, differing between subjects treated with prednisone alone and those receiving thymectomy plus prednisone. These findings provide unique insights into the biological effects of thymectomy. Notably, application of multiple machine learning approaches revealed non-linear relationships in the identification of outcome-predictive proteins in the ETTX plus prednisone group, while independent linear (additive) patterns predominated among those receiving prednisone alone. The proteins associated with treatment response were primarily involved in T cell signaling and cell trafficking.\u003c/p\u003e\u003cp\u003eThe limitations of our work lie in those inherent to a rare disease research of small sample size, and despite our machine learning nested cross validation approach, future validation studies will be required, including those that assess change in the proteome over time. Further, the benefit of a rigorous clinical trial data set compromises the potential for immediate applicability to a real-world population with a greater diversity of patients with multiple co-morbidities. This includes patients outside the age of study entrance criteria and AChR antibody status.\u003c/p\u003e\u003cdiv id=\"Sec24\" class=\"Section2\"\u003e\u003ch2\u003eMyasthenia gravis driven differences from controls\u003c/h2\u003e\u003cp\u003eOur initial hypothesis was that prednisone treatment would significantly alter the serum proteome and potentially confound disease-specific differences; however, this was not observed. Only 22 DEPs were identified between subjects on prednisone and those not receiving treatment at study entry, compared to over 400 DEPs observed between MG subjects and healthy controls, regardless of prednisone use (Supplemental Fig.\u0026nbsp;1). These findings align with those of Nelke and colleagues, who reported that, in a highly heterogeneous treated cohort of MG patients, proteomic differences were greater between patients and a small control group (n\u0026thinsp;=\u0026thinsp;10) than among the four patient consensus clusters (\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e). Despite the widespread and long-standing use of corticosteroids across a range of diseases, surprisingly little is known about their effects on gene and protein expression following chronic administration. Serum proteomic changes are presumed to result from the genomic actions of corticosteroids, which are themselves modulated by physiological adaptations, including downregulation of the glucocorticoid receptor and feedback regulation of downstream pathways (\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e). Moreover, glucocorticoid effects on tissue may not be reflected in the circulating proteome, as our findings suggest.\u003c/p\u003e\u003cp\u003eThe proteomic signatures differences were a function of MG. Pathway analyses revealed alterations in cytoskeletal and organelle organization. The explanation for such changes in circulation may lie in injury to the large number of neuromuscular junctions across the body. AChR antibodies induce complement-mediated injury and deposits of immunoglobulins, complement components, membrane-attack complex, and AChR are found in the synaptic cleft. Macrophage infiltration is not appreciated and therefore, these components and others could make their way into circulation. Complement components of the membrane attack complex are appreciated to be shed with tissue trauma, as would occur in MG, and these could have a pro- or anti-inflammatory effect (\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e). The same can be said for mechanisms of modulating effects on lymphocytes. Immune cells must migrate from blood to various lymphoid organs and other tissues. KEGG pathway analysis identified leukocyte transendothelial migration, B cell receptor signaling pathway, and Fc receptor pathways to be reduced in the MG subjects (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB). Similarly, GSEA evaluation of the GO-BP database was supportive of these results, but also found immunoglobulin production and complement activation were significantly downregulated in MG patients compared to controls. Complement-mediation of disease pathology of MG is well appreciated (\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e). T cell receptor signaling pathway was significantly upregulated (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eD). These results may seem counterintuitive that these pathways would be reduced in an active autoimmune disorder; however, this would represent a mechanism to suppress global immune activity in the context of the autoimmune reaction. T cell receptor signaling pathway was significantly upregulated, which could be a suppressor activity. Similar pathways were identified in our transcriptional profiling analysis of thymus obtained from the MGTX study (\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eConsistent with pathway analysis, the most significantly expressed DEPs were related to cytoskeletal function or directly to immune function with several overlapping. TAGLN2, PDLIM2, CNN2, and SEPTIN6 were all originally characterized by cytoskeletal association; their involvement in T cell activation and inflammatory signaling (\u003cspan additionalcitationids=\"CR39\" citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e) make them particularly interesting proteins. CSK, or c-Src terminal kinase, is involved in T cell and B cell receptor signaling and can be found in exosomes (\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e). Many of the proteins identified are highly likely to be carried in exosomes, which are known to be secreted by a number of cell types in response to stress signals to promote tissue healing (\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e).\u003c/p\u003e\u003cdiv id=\"Sec25\" class=\"Section3\"\u003e\u003ch2\u003eMachine Learning Assessment\u003c/h2\u003e\u003cp\u003eTo robustly capture relationships between serum proteins and clinical improvement, we applied a panel of machine learning algorithms, each with distinct pattern sensitivities. First, we used Boruta analysis to identify features that \u003cem\u003ecollectively\u003c/em\u003e contributed to the prediction of treatment response. In our study, features are proteins. Proteins that do not improve the model\u0026rsquo;s predictive performance are not selected. Importantly, proteins that may be part of the same biological pathway might not all be selected if their inclusion would not provide additional predictive value beyond what is already captured by other selected proteins. Proteins identified through feature selection were utilized in several machine learning algorithms to assess the AUC of the proteins predicting outcomes. This assumption-agnostic approach, combined with nested cross-validation, enabled head-to-head model comparison and revealed Random Forests as optimal for PA-treated patients and XGBoost for those receiving ETTX; each performed excellently with an AUC\u0026thinsp;\u0026gt;\u0026thinsp;0.9.\u003c/p\u003e\u003cp\u003eRandom Forests and XGBoost often outperform probabilistic classifiers (Na\u0026iuml;ve Bayes, Logistic Regression) and regularized linear models (LASSO, Elastic Net) in complex biological datasets, including large proteomic data sets. This finding lies in the nonlinearity of proteomic data which is further complicated by the correlation of many proteins. XGBoost, is also tree-based ensembles that use boosting. Unlike Random Forests, where trees are built independently, boosting builds trees sequentially, with each new tree focusing on correcting the mistakes made by the previous ones. This allows the model to gradually improve its performance by learning from its own errors. XGBoost builds on this idea by adding regularization to penalize overly complex models, pruning to eliminate unhelpful splits, and parallelization to speed up training.\u003c/p\u003e\u003cp\u003eThe eight models represent a spectrum of complexity, underlying model assumptions, and learning strategies, which allowed us to systematically evaluate predictive performance across our proteomic datasets. The use of multiple ML algorithms is critical in biomarker discovery, where signals are often heterogeneous and no single method can guarantee optimal performance (\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e, \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e). The multi-model approach not only increases confidence that our findings were not artifacts of any one modeling assumption and provides insight into the nature of the predictive structure. Overall, these differences highlight how proteomic landscapes shape clinical response in a treatment-dependent manner, with additive effects characterizing PA subjects and nonlinear interactions driving improvement in those receiving ETTX.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec26\" class=\"Section3\"\u003e\u003ch2\u003ePrevious proteomic evaluations in myasthenia gravis\u003c/h2\u003e\u003cp\u003eOur study is among the first to utilize both proteomics and ML modeling to predict short-term treatment response in patients with MG. The number of studies utilizing machine learning on biospecimen datasets, not just clinical data, from patients with MG without thymoma is sparse. This includes studies identifying biomarkers from the gut microbiome of MG using metabolomics (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e) and rRNA sequencing (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e), from an MG monozygotic twin microarray dataset (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e), and from proteomic data from AChR-ab positive MG patients (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eA recent study by Lin et al also utilized LC-MS DIA proteomics to assess pre-surgical differences in serum between effective and ineffective thymectomy in MG; however, samples were from a cohort of patients with thymoma MG (\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e). Even so, they reported that the protein SELL was downregulated in treatment-responsive thymectomized patients at baseline. This finding is partially in line with our study, in which SELL was considered an important protein in predicting treatment response specifically in patients receiving a thymectomy and prednisone (ETTX patients), albeit the relative expression in ETTX patients who improved was higher than those that did not (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). SELL (L-selectin) has been strongly implicated for its role in the adhesion and trans endothelial migration of leukocytes, especially neutrophils, before being cleared via ectodomain shedding, acting as a marker for neutrophil activation (\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eFive studies used targeted (i.e., O-link) proteomics to investigate changes between MG and HC samples (\u003cspan additionalcitationids=\"CR50 CR51 CR52\" citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e). In agreement with results reported by Bhandage et al, 2024 and Molin et al, 2017, we found elevated levels of the pro-inflammatory calcium binding protein SA100-A12 in MG patients compared to controls, which may be indicative of increased neutrophil activation and lymphocyte recruitment (\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e, \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003cem\u003eProteomic Signature of Prednisone Treatment Effect.\u003c/em\u003e\u003c/p\u003e\u003cp\u003eDEPs at baseline were identified that were associated with improved QMG after 6 months and these were distinct for ETTX and PA groups, despite the well-matched clinical characteristics of these subjects. Seven proteins (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) were associated with treatment response, with CCL16 and IGHV3-43 being the most clearly related to the immune response. CCL16 is a chemokine primarily expressed in the liver but also the thymus and some lymphocytes. CCL16 is appreciated as being responsible for recruitment of T cells to sites of pathology and other chemokines are well-established to be involved in MG, this has not been appreciated before for CCL16 (\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e). IGHV3-43 is a portion of the variable region of the heavy chain of IgG, which would be expected to be detected in serum, but why the lower level of this specific IgG heavy chain in serum would be associated with improvement is not clear. However, given that genetic variations in the immunoglobulin heavy chain locus may drive antibody response could suggest this IGHV3-43 may be of particular importance in MG (\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e). Results from the GeneAgent analysis similarly identified the immune response roles, and a mild predominance of ribosomal involvement by RPS3 and RPS3A.\u003c/p\u003e\u003cp\u003e\u003cem\u003eProteomic Signature for Thymectomy plus Prednisone Treatment Effect.\u003c/em\u003e\u003c/p\u003e\u003cp\u003eThe 17 proteins identified by Boruta analysis as predictive of clinical improvement in MG patients treated with thymectomy plus prednisone likely reflect a coordinated biological response to immune modulation and tissue remodeling. These proteins encompass key processes including immune regulation and complement control (CD55, SELL, IGKV1-5, RNASET2), extracellular matrix and stromal reorganization (MMP3, FBLN5, CCN5), cytoskeletal dynamics and cellular trafficking (ARHGAP9, RAB14, RDX), stress response and protein repair (HSP90AA1, PCMT1), and modulation of the BMP/TGF-β axis involved in thymic epithelial architecture (CHRDL1, SPP2). Collectively, these proteins may serve as systemic indicators of thymic regression, loss of germinal center activity, suppression of autoreactive B and T cell signaling, and the resolution of inflammatory and structural abnormalities that drive disease activity. Their predictive value likely lies in their ability to capture early molecular changes that precede and enable long-term clinical benefit as observed in the MGTX trial. From a translational standpoint, the biological roles of these proteins suggest potential therapeutic targets or mechanistic entry points for understanding how thymectomy induces systemic immune reprogramming.\u003c/p\u003e\u003cp\u003eSVEP1 had the greatest SHAP value predictive of improvement (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eD, E). SVEP1 (Sushi, von Willebrand factor type A) is a large extracellular matrix protein that has been identified to be associated with poor outcomes in several conditions including coronary artery disease, dementia, and hypertension and with poor outcomes in heart failure, where it is strongly associated with activated T cells (\u003cspan additionalcitationids=\"CR59\" citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e). SVEP1 is composed primarily of complement control protein (CCP, also known as Sushi) domains; whether it binds complement proteins is not known. Elevated SVEP1 levels in blood are also associated with diseases of aging, and SVEP1 has been shown to activate the mTOR pathway, a key regulator of longevity. Given autoimmune disorders increase with chronological age, these findings provide the intriguing insight that these early-onset MG patients have an advanced biological age.\u003c/p\u003e\u003cp\u003eHeat shock protein 90α (Hsp90α) is a ubiquitously expressed heat shock protein, which interacts with close to two thousand proteins and numerous biological pathways (\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e, \u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e). Elevations in serum are associated with greater cancer risk and vascular inflammation. Hsp90α has been shown to suppress or activate immune processes depending on specific conditions. SPP2 is produced in the liver and known to be found in circulation. It binds and activates intracellular signaling pathways of the TGF-beta superfamily cytokines. Interestingly, SPP2 shares sequence homology with osteopontin. Osteopontin gene polymorphisms are associated with treatment response to prednisone in MG (\u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e63\u003c/span\u003e). CD55 is a complement inhibitory protein known to be concentrated at the neuromuscular junction and protective of injury to antibody-mediated, complement-induced in experimental MG (\u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e). While typically found on cell surfaces, CD55 exists in soluble form and in culture reduces proliferation peripheral blood monocytes and would be expected to modulate complement activity (\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e, \u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e66\u003c/span\u003e). CD55 has four CCP domains and therefore shares structural similarities to SVEP1.\u003c/p\u003e\u003cp\u003eThe thymus of patients with MG caused by AChR antibodies demonstrates thymic hyperplasia characterized by lymphoid follicles and germinal center formation, which was documented for the present group of patients (\u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e). The germinal centers are a source of AChR-specific B cells and contain autoreactive T cells. Removal of the thymus reduces a source of autoreactive T and B cells but, despite its removal, clinical disease persists in a variable number of patients.\u003c/p\u003e\u003cp\u003eThe panel of proteins predictive of outcome in ETTX patients supports a fundamental difference in individual susceptibility to removal of thymus. The proteins identified and our approach to this question reveals that a single factor, such as T regulatory cell dysfunction, as appreciated in patients with MG, is unlikely to be the single answer to treatment effect of thymectomy. GeneAgent results supported critical roles of cell trafficking in these proteins; this would be consistent with the susceptibility of such pathways to treatment outcome based on the exodus of pathogenic cells from the thymus to peripheral immune organs.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec27\" class=\"Section3\"\u003e\u003ch2\u003eClinical Impact\u003c/h2\u003e\u003cp\u003eThe clinical importance of our findings lies in the identification of serum protein signatures associated with treatment response to thymectomy plus prednisone and prednisone alone, offering a potential tool for personalized treatment planning in myasthenia gravis. Currently, decisions regarding thymectomy are based on generalized criteria without reliable biomarkers to predict individual benefit (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). Our results suggest that proteomic profiling could enable clinicians to identify patients more likely to respond favorably to thymectomy, thereby reducing unnecessary surgeries in non-responders, expediting treatment decisions for those likely to benefit, and guiding others toward alternative therapies. The observed heterogeneity in treatment response is consistent with findings by Nelke et al., who identified patient subgroups with differential responsiveness, including one enriched for complement inhibition benefit (\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e). Further, the prednisone-only related signature would identify poor responders to the standard of care treatment for MG (\u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e). This approach parallels precision oncology strategies, where surgical or pharmacologic interventions are tailored using molecular markers. Moreover, stratification based on this proteomic signature could inform the development of novel therapeutic targets for non-responders.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"Conclusions","content":"\u003cp\u003eBaseline serum proteomics distinguished treatment-specific predictors of clinical response in patients with MG. PA-patients\u0026rsquo; response was explained mainly by single protein markers, whereas ETTX-patients\u0026rsquo; response was predicted by complex protein interactions. These findings represent an important first step toward biomarker-guided treatment in MG; with validation in independent cohorts, proteomic signatures could refine patient stratification, inform clinical trial design, and enable more personalized therapeutic decisions.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eMG: Myasthenia gravis\u003c/p\u003e\n\u003cp\u003eMGTX: Thymectomy Trial in Non-Thymomatous Myasthenia Gravis Patients Receiving Prednisone Therapy\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAChR: acetylcholine receptor\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eETTX: thymectomy plus prednisone\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePA: prednisone alone\u003c/p\u003e\n\u003cp\u003eQMG: Quantitative MG\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eLC-MS: liquid chromatography-mass spectrometry\u003c/p\u003e\n\u003cp\u003eML: machine learning\u003c/p\u003e\n\u003cp\u003eIRB: institutional review board\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDDA: data dependent acquisition\u003c/p\u003e\n\u003cp\u003eDIA: data independent acquisition\u003c/p\u003e\n\u003cp\u003eDEP: differentially expressed protein\u003c/p\u003e\n\u003cp\u003eHC: healthy control\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eORA: over representation analysis\u003c/p\u003e\n\u003cp\u003eGSEA: gene set enrichment analysis\u003c/p\u003e\n\u003cp\u003enCV: nested cross validation\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eLASSO: least absolute shrinkage and selection operator\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCART: classification and regression tree algorithms\u003c/p\u003e\n\u003cp\u003eAdaBoost, Ada: adaptive boosting\u003c/p\u003e\n\u003cp\u003eXGBoost: extreme gradient boosting\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePCA: principal component analysis\u003c/p\u003e\n\u003cp\u003eGO-BP: gene ontology biological processes\u003c/p\u003e\n\u003cp\u003eKEGG: Kyoto Encyclopedia of Genes and Genomes\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSEM: standard error mean\u003c/p\u003e\n\u003cp\u003eNB: Na\u0026iuml;ve Bayes\u003c/p\u003e\n\u003cp\u003eEN: elastic net\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCT: classification tree\u003c/p\u003e\n\u003cp\u003eRF: random forest\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAUC: area under the ROC curve\u003c/p\u003e\n\u003cp\u003eROC: receiver operating characteristic\u003c/p\u003e\n\u003cp\u003eAUCPR: area under the precision-recall curve\u003c/p\u003e\n\u003cp\u003eMCC: Matthew\u0026rsquo;s correlation coefficient\u003c/p\u003e\n\u003cp\u003eSHAP: Shapley additive explanations\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCCL16: C-C motif chemokine ligand 16\u003c/p\u003e\n\u003cp\u003eCILP2: cartilage intermediate layer protein 2\u003c/p\u003e\n\u003cp\u003eGALNT1: polypeptide N-acetylgalactosaminyltransferase 1\u003c/p\u003e\n\u003cp\u003eIGHV3-43: immunoglobulin heavy variable 3-43\u003c/p\u003e\n\u003cp\u003eRPS3: ribosomal protein S3\u003c/p\u003e\n\u003cp\u003eRPS3A: ribosomal protein S3A\u003c/p\u003e\n\u003cp\u003eVAPB: VAMP (vesicle-associated membrane protein)-associated protein B and C\u003c/p\u003e\n\u003cp\u003eARHGAP9: Rho GTPase activating protein 9\u003c/p\u003e\n\u003cp\u003eCCN5: cellular communication network factor 5\u003c/p\u003e\n\u003cp\u003eCD55: decay accelerating factor for complement\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCHAD: chondroadherin\u003c/p\u003e\n\u003cp\u003eCHRDL1: chordin like 1\u003c/p\u003e\n\u003cp\u003eFBLN5: fibulin 5\u003c/p\u003e\n\u003cp\u003eHS1BP3: HCLS1 binding protein 3\u003c/p\u003e\n\u003cp\u003eHSP90AA1: heat shock protein 90 alpha family class A member 1\u003c/p\u003e\n\u003cp\u003eIGKV1-5: immunoglobulin kappa variable 1-5\u003c/p\u003e\n\u003cp\u003eMMP3: matrix metallopeptidase 3\u003c/p\u003e\n\u003cp\u003ePCMT1: protein-L-isoaspartate (D-aspartate) O-methyltransferase\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eRAB14: member RAS oncogene family, F protein-binding protein 1\u003c/p\u003e\n\u003cp\u003eRDX: radixin\u003c/p\u003e\n\u003cp\u003eRNASET2: ribonuclease T2\u003c/p\u003e\n\u003cp\u003eSELL: selectin L, lymphocyte adhesion molecule 1\u003c/p\u003e\n\u003cp\u003eSPP2: secreted phosphoprotein 2\u003c/p\u003e\n\u003cp\u003eSVEP1: sushi, von Willebrand factor A, EGF and pentraxin domain containing 1\u0026nbsp;\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cem\u003eEthics approval and consent to participate\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe NINDS funded the trial and assembled an independent Data Safety Monitoring Board. Sites received local institutional review board (IRB)/ethics committee approvals, and each patient provided written informed consent before study entry including provision of serum samples. All specimens were deidentified. The George Washington University IRB provided additional review and approved these investigations.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eConsent for publication\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAvailability of data and materials\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe R code used to run nCV is available on GitHub at https://github.com/drkgil/NestCV. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe datasets used and analyzed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eCompeting interests\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAll financial and non-financial competing interests must be declared in this section\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eKG: none.\u003c/p\u003e\n\u003cp\u003eAC: none.\u003c/p\u003e\n\u003cp\u003eHK:\u0026nbsp;receives consulting fees from Novartis, UCB Pharmaceuticals, Kite Pharmaceuticals, Canopy Immunotherapeutics, and Merck. Argenx provides an unrestricted educational grant to George Washington University. He is an unpaid consultant for Care Constitution. HK has equity interest in Mimivax, LLC. HK is principal investigator for the Rare Disease Network, MGNet supported by NIH grant U54NS115054.\u003c/p\u003e\n\u003cp\u003eLK: receives consulting fees from Novartis. LK has equity interest in Mimivax LLC. LK is a co-investigator for the Rare Disease Network, MGNet supported by NIH grant U54NS115054.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eFunding\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe work was supported by the MGNet, a member of the Rare Disease Clinical Research Network Consortium (RDCRN) NIH U54 NS115054 and NIH/NINDS U01 NS042685 (MGTX). Additional support was provided by philanthropic support for the Laboratory for MG Research at George Washington University. Funding support for the Data Management and Coordinating Center (DMCC) is provided by the National Center for Advancing Translational Sciences (NCATS) and the National Institute of Neurological Disorders and Stroke (NINDS).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAuthors\u0026rsquo; contributions\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eKG: Data curation, Formal analysis, Software, Writing \u0026ndash; original draft, Writing \u0026ndash; review \u0026amp; editing.\u003cbr\u003e\u0026nbsp;AKC: Data interpretation, Writing \u0026ndash; review \u0026amp; editing.\u003cbr\u003e\u0026nbsp;HK: Conceptualization, Data curation, Funding acquisition, Resources, Supervision, Writing \u0026ndash; original draft, Writing \u0026ndash; review \u0026amp; editing.\u003cbr\u003e\u0026nbsp;LK: Conceptualization, Formal analysis, Funding acquisition, Methodology, Supervision, Writing \u0026ndash; review \u0026amp; editing.\u003c/p\u003e\n\u003cp\u003eAll authors read and approved the final manuscript.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAcknowledgements\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eLC-MS performed by the Core Services at Lombardi Comprehensive Cancer Center, Georgetown University Medical Center, Washington, DC, 20057, USA. We would also like to thank the Metabolomics Shared Resource, partially supported by NIH/NCI grant P30-CA051008.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eKaminski HJ, Sikorski P, Coronel SI, Kusner LL. Myasthenia gravis: the future is here. J Clin Invest 2024 June 17;134(12):e179742.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBenatar M, Cutter G, Kaminski HJ. The best and worst of times in therapy development for myasthenia gravis. Muscle Nerve. 2023;67(1):12\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWolfe GI, Kaminski HJ, Aban IB, Minisman G, Kuo HC, Marx A, et al. Randomized Trial of Thymectomy in Myasthenia Gravis. N Engl J Med. 2016;375(6):511\u0026ndash;22.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAlhaidar MK, Abumurad S, Soliven B, Rezania K. Current Treatment of Myasthenia Gravis. J Clin Med. 2022;11(6):1597.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVerschuuren JJ, Palace J, Murai H, Tannemaat MR, Kaminski HJ, Bril V. Advances and ongoing research in the treatment of autoimmune neuromuscular junction disorders. Lancet Neurol. 2022;21(2):189\u0026ndash;202.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFahrner LJ, Chen E, Topol E, Rajpurkar P. The generative era of medical AI. Cell 2025 July 10;188(14):3648\u0026ndash;60.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXu C, Jackson SA. Machine learning and complex biological data. Genome Biol. 2019;20(1):76.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJin S, Yin J, Li W, Mao N. Mitochondrial autophagy-related gene signatures associated with myasthenia gravis diagnosis and immunity. Autoimmunity. 2025;58(1):2465410.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu H, Liu G, Guo R, Li S, Chang T. Identification of Potential Key Genes for the Comorbidity of Myasthenia Gravis With Thymoma by Integrated Bioinformatics Analysis and Machine Learning. Bioinforma Biol Insights. 2024;18:11779322241281652.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhou G, Wang S, Lin L, Lu K, Lin Z, Zhang Z, et al. Screening for immune-related biomarkers associated with myasthenia gravis and dilated cardiomyopathy based on bioinformatics analysis and machine learning. Heliyon. 2024;10(7):e28446.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSchroeter CB, Nelke C, Stascheit F, Huntemann N, Preusse C, Dobelmann V et al. Inter-alpha-trypsin inhibitor heavy chain H3 is a potential biomarker for disease activity in myasthenia gravis. Acta Neuropathol (Berl). 2024 June 18;147(1):102.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChang CC, Liu TC, Lu CJ, Chiu HC, Lin WN. Explainable machine learning model for identifying key gut microbes and metabolites biomarkers associated with myasthenia gravis. Comput Struct Biotechnol J. 2024;23:1572\u0026ndash;83.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChang CC, Liu TC, Lu CJ, Chiu HC, Lin WN. Machine learning strategy for identifying altered gut microbiomes for diagnostic screening in myasthenia gravis. Front Microbiol [Internet]. 2023 Sept 27 [cited 2025 July 15];14. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.frontiersin.org/journals/microbiology/articles/\u003c/span\u003e\u003cspan address=\"https://www.frontiersin.org/journals/microbiology/articles/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fmicb.2023.1227300/full\u003c/span\u003e\u003cspan address=\"10.3389/fmicb.2023.1227300/full\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBauman TA, Lee SM, Juel VC, Li Y, Gilbert K, Chen J, et al. Serum fibrinogen is not elevated in patients with myasthenia gravis. Sci Rep. 2025;15(1):13013.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCheema AK, Li Y, Ventimiglia M, Kowalczyk K, Hankins R, Bandi G, et al. Radiotherapy Induces Innate Immune Responses in Patients Treated for Prostate Cancers. Clin Cancer Res Off J Am Assoc Cancer Res. 2023;29(5):921\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHsiao Y, Zhang H, Li GX, Deng Y, Yu F, Valipour Kahrood H, et al. Analysis and Visualization of Quantitative Proteomics Data Using FragPipe-Analyst. J Proteome Res. 2024;23(10):4303\u0026ndash;15.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePhipson B, Lee S, Majewski IJ, Alexander WS, Smyth GK. Robust hyperparameter estimation protects against hypervariable genes and improves power to detect differential expression. Ann Appl Stat. 2016 June;10(2):946\u0026ndash;63.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRitchie ME, Phipson B, Wu D, Hu Y, Law CW, Shi W, et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 2015;43(7):e47.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYu G. Thirteen years of clusterProfiler. Innov. 2024;5(6):100722.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYu G, Wang LG, Han Y, He QY. clusterProfiler: an R Package for Comparing Biological Themes Among Gene Clusters. OMICS J Integr Biol. 2012;16(5):284\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePudjihartono N, Fadason T, Kempa-Liehr AW, O\u0026rsquo;Sullivan JM. A Review of Feature Selection Methods for Machine Learning-Based Disease Risk Prediction. Front Bioinforma. 2022 June;27:2:927312.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKursa MB, Rudnicki WR. Feature Selection with the Boruta Package. J Stat Softw 2010 Sept 16;36:1\u0026ndash;13.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMeyer D, Dimitriadou E, Hornik K, Weingessel A, Leisch F, Chang CC et al. e1071: Misc Functions of the Department of Statistics, Probability Theory Group (Formerly: E1071), TU Wien [Internet]. 2024 [cited 2025 July 28]. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://cran.r-project.org/web/packages/e1071/index.html\u003c/span\u003e\u003cspan address=\"https://cran.r-project.org/web/packages/e1071/index.html\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFriedman J, Tibshirani R, Hastie T. Regularization Paths for Generalized Linear Models via Coordinate Descent. J Stat Softw. 2010;33(1):1\u0026ndash;22.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTay JK, Narasimhan B, Hastie T. Elastic Net Regularization Paths for All Generalized Linear Models. J Stat Softw. 2023;106(1):1\u0026ndash;31.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTherneau T, Atkinson B, rpart. Recursive Partitioning and Regression Trees [Internet]. 2025. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://CRAN.R-project.org/package=rpart\u003c/span\u003e\u003cspan address=\"https://CRAN.R-project.org/package=rpart\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiaw A, Wiener M. Classification and Regression by randomForest. R News. 2002;2(3):18\u0026ndash;22.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOlson M, JOUSBoost. Implements Under/Oversampling for Probability Estimation [Internet]. 2017. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://CRAN.R-project.org/package=JOUSBoost\u003c/span\u003e\u003cspan address=\"https://CRAN.R-project.org/package=JOUSBoost\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen T, Guestrin C, XGBoost:. A Scalable Tree Boosting System. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining [Internet]. New York, NY, USA: Association for Computing Machinery; 2016 [cited 2025 Sept 16]. pp. 785\u0026ndash;94. (KDD \u0026rsquo;16). Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://dl.acm.org/doi/10.1145/2939672.2939785\u003c/span\u003e\u003cspan address=\"https://dl.acm.doi/10.1145/2939672.2939785\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang Z, Jin Q, Wei CH, Tian S, Lai PT, Zhu Q, et al. GeneAgent: self-verification language agent for gene-set analysis using domain databases. Nat Methods. 2025;22(8):1677\u0026ndash;85.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBarohn RJ, McIntire D, Herbelin L, Wolfe GI, Nations S, Bryan WW. Reliability testing of the quantitative myasthenia gravis score. Ann N Y Acad Sci. 1998;841:769\u0026ndash;72.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNelke C, Schroeter CB, Barman S, Stascheit F, Masanneck L, Theissen L, et al. Identification of disease phenotypes in acetylcholine receptor-antibody myasthenia gravis using proteomics-based consensus clustering. EBioMedicine. 2024 July;105:105231.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAyyar VS, Jusko WJ. Transitioning from Basic toward Systems Pharmacodynamic Models: Lessons from Corticosteroids. Pharmacol Rev. 2020;72(2):414\u0026ndash;38.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKarasu E, Eisenhardt SU, Harant J, Huber-Lang M. Extracellular Vesicles: Packages Sent With Complement. Front Immunol. 2018;9:721.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAlbazli K, Kaminski HJ, Howard JF. Complement Inhibitor Therapy for Myasthenia Gravis. Front Immunol. 2020;11:917.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJacob S, Howard JF. Neuromuscular junction and the complement system. Int Rev Neurobiol. 2025;182:21\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSengupta M, Wang BD, Lee NH, Marx A, Kusner LL, Kaminski HJ. MicroRNA and mRNA expression associated with ectopic germinal centers in thymus of myasthenia gravis. PLoS ONE. 2018;13(10):e0205464.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eIvanov AI, Le HT, Naydenov NG, Rieder F. Novel Functions of the Septin Cytoskeleton: Shaping Up Tissue Inflammation and Fibrosis. Am J Pathol. 2021;191(1):40\u0026ndash;51.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShi Y, Fan T, Yang Y, Liu J, Ouyang J, Dai J. Beyond structural domains: the emerging roles of PDLIM2 in cellular signaling and cancer progression. Front Physiol. 2025;16:1569285.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYin LM, Ulloa L, Yang YQ. Transgelin-2: Biochemical and Clinical Implications in Cancer and Asthma. Trends Biochem Sci. 2019;44(10):885\u0026ndash;96.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHikita T, Kuwahara A, Watanabe R, Miyata M, Oneyama C. Src in endosomal membranes promotes exosome secretion and tumor progression. Sci Rep. 2019;9(1):3265.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhu S, Wang H, Ranjan K, Zhang D. Regulation, targets and functions of CSK. Front Cell Dev Biol. 2023;11:1206539.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAoi W, Tanimura Y. Roles of Skeletal Muscle-Derived Exosomes in Organ Metabolic and Immunological Communication. Front Endocrinol. 2021;12:697204.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDiaz-Uriarte R, G\u0026oacute;mez de Lope E, Giugno R, Fr\u0026ouml;hlich H, Nazarov PV, Nepomuceno-Chamorro IA, et al. Ten quick tips for biomarker discovery and validation analyses using machine learning. PLoS Comput Biol. 2022;18(8):e1010357.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang X, Jonassen I, Goks\u0026oslash;yr A. Machine Learning Approaches for Biomarker Discovery Using Gene Expression Data. In: Helder I. N, editor. Bioinformatics [Internet]. Brisbane (AU): Exon Publications; 2021 [cited 2025 July 24]. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.ncbi.nlm.nih.gov/books/NBK569564/\u003c/span\u003e\u003cspan address=\"http://www.ncbi.nlm.nih.gov/books/NBK569564/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLin X, Qi G, Liu P. Identification of potential prognostic biomarkers of thymoma with myasthenia gravis based on serum proteomics. Front Immunol. 2025;16:1580219.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eIvetic A, Hoskins Green HL, Hart SJ. L-selectin: A Major Regulator of Leukocyte Adhesion, Migration and Signaling. Front Immunol. 2019;10:1068.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRahman I, Collado S\u0026aacute;nchez A, Davies J, Rzeniewicz K, Abukscem S, Joachim J, et al. L-selectin regulates human neutrophil transendothelial migration. J Cell Sci. 2021;134(3):jcs250340.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBhandage AK, Kenina V, Huang YF, Roddate M, Kauke G, Grosmane A, et al. Serum protein biomarker profile distinguishes acetylcholine receptor antibody seropositive myasthenia gravis patients from healthy controls. iScience. 2024;27(8):110564.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBhandage AK, Punga AR. Inflammatory imbalance and activation deficits in T cells of myasthenia gravis patients revealed by proteomic profiling. Front Immunol [Internet]. 2025 Aug 4 [cited 2025 Aug 11];16. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.frontiersin.org/journals/immunology/articles/\u003c/span\u003e\u003cspan address=\"https://www.frontiersin.org/journals/immunology/articles/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fimmu.2025.1648020/full\u003c/span\u003e\u003cspan address=\"10.3389/fimmu.2025.1648020/full\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMolin CJ, Westerberg E, Punga AR. Profile of upregulated inflammatory proteins in sera of Myasthenia Gravis patients. Sci Rep. 2017;7:39716.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMyllynen C, Sarkkinen J, Atula S, Tienari P, Kek\u0026auml;l\u0026auml;inen E, Laakso SM. A skewed ratio of free light chains is more common in patients with late-onset than early-onset myasthenia gravis. Immunol Lett. 2023;260:81\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOkuzono Y, Miyakawa S, Itou T, Sagara M, Iwata M, Ishizuchi K, et al. B-cell immune dysregulation with low soluble CD22 levels in refractory seronegative myasthenia gravis. Front Immunol. 2024;15:1382320.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBobek D, Sestan M, Mijacika L, Kovacic N, Lukic IK, Grcevic D, et al. Serum S100A12 levels in children with childhood-onset systemic lupus erythematosus, systemic juvenile arthritis, and systemic undefined recurrent fevers. Z Rheumatol. 2023;82(Suppl 1):74\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDonato R, Cannon BR, Sorci G, Riuzzi F, Hsu K, Weber DJ, et al. Functions of S100 Proteins. Curr Mol Med. 2013;13(1):24\u0026ndash;57.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSun L, Su Y, Jiao A, Wang X, Zhang B. T cells in health and disease. Signal Transduct Target Ther. 2023 June;19(1):235.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRodriguez OL, Safonova Y, Silver CA, Shields K, Gibson WS, Kos JT et al. Genetic variation in the immunoglobulin heavy chain locus shapes the human antibody repertoire. Nat Commun. 2023 July 21;14(1):4419.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eElenbaas JS, Jung IH, Coler-Reilly A, Lee PC, Alisio A, Stitziel NO. The emerging Janus face of SVEP1 in development and disease. Trends Mol Med. 2023;29(11):939\u0026ndash;50.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi H, Huang C, Su R, Wang M, Ma Y, Wang Y, et al. Developing a Panel of Shared Susceptibility Genes as Diagnostic Biomarkers for chronic obstructive pulmonary disease and Heart Failure. Comput Biol Med. 2025 Sept;196:110657.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLotfinaghsh A, Imam A, Pompian A, Stitziel NO, Javaheri A. Clinical Insights from Proteomics in Heart Failure. Curr Heart Fail Rep. 2025;22(1):12.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHu C, Yang J, Qi Z, Wu H, Wang B, Zou F, et al. Heat shock proteins: Biological functions, pathological roles, and therapeutic opportunities. MedComm. 2022;3(3):e161.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZuehlke AD, Beebe K, Neckers L, Prince T. Regulation and function of the human HSP90AA1 gene. Gene. 2015;570(1):8\u0026ndash;16.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXie Y, Li HF, Sun L, Kusner LL, Wang S, Meng Y, et al. The Role of Osteopontin and Its Gene on Glucocorticoid Response in Myasthenia Gravis. Front Neurol. 2017;8:230.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKusner LL, Kaminski HJ, Soltys J. Effect of complement and its regulation on myasthenia gravis pathogenesis. Expert Rev Clin Immunol. 2008;4(1):43\u0026ndash;52.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKwon YC, Kim H, Meyer K, Di Bisceglie AM, Ray R. Distinct CD55 Isoform Synthesis and Inhibition of Complement-Dependent Cytolysis by Hepatitis C Virus. J Immunol Baltim Md 1950. 2016;197(4):1127\u0026ndash;36.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSio YY, Anantharaman R, Lee SQE, Matta SA, Ng YT, Chew FT. The Asthma-associated PER1-like domain-containing protein 1 (PERLD1) Haplotype Influences Soluble Glycosylphosphatidylinositol Anchor Protein (sGPI-AP) Levels in Serum and Immune Cell Proliferation. Sci Rep. 2020;10(1):715.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWeis CA, Aban IB, Cutter G, Kaminski HJ, Scharff C, Grie\u0026szlig;mann BW, et al. Histopathology of thymectomy specimens from the MGTX-trial: Entropy analysis as strategy to quantify spatial heterogeneity of lymphoid follicle and fat distribution. PLoS ONE. 2018;13(6):e0197435.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKaminski HJ, Denk J. Corticosteroid Treatment-Resistance in Myasthenia Gravis. Front Neurol. 2022;13:886625.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"journal-of-translational-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jtrm","sideBox":"Learn more about [Journal of Translational Medicine](http://translational-medicine.biomedcentral.com)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/jtrm/default.aspx","title":"Journal of Translational Medicine","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"myasthenia gravis, machine learning, treatment response, proteomics, mass spectrometry, prednisone, thymectomy","lastPublishedDoi":"10.21203/rs.3.rs-7915342/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7915342/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eMyasthenia gravis (MG) is a prototypical antibody-mediated autoimmune disease with variable treatment responses with a need for biomarkers to guide therapeutic decision making. Proteomic profiling, coupled with machine learning, offers a powerful approach to identify biomarkers that may predict treatment response.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eWe analyzed sera collected at entry (baseline) from participants in a phase 3 trial randomized trial comparing thymectomy plus prednisone versus prednisone alone, along with matched controls using liquid chromatography\u0026ndash;mass spectrometry. We derived disease-specific proteomic signatures and evaluated associations between baseline proteins and 6-month clinical outcomes using multiple machine-learning approaches with internal validation.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eBaseline serum proteomes distinguished MG from controls, with pathway enrichment implicating complement activation, immunoglobulin production, and T-cell receptor signaling. Distinct protein panels predicted 6-month clinical improvement within each treatment arm. In the thymectomy-plus-prednisone group, models captured non-linear relationships of predictive proteins in contrast with the predominant additive patterns observed in the prednisone-alone group. Predictive proteins were enriched for T-cell signaling and leukocyte trafficking functions, providing insight into treatment-specific biology.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eBaseline serum proteomics captures core disease characteristics of MG and predicts short-term clinical response in a treatment-specific manner. If validated in independent cohorts, these findings could enable biomarker-guided selection of thymectomy, refine risk stratification, and furnish mechanistic readouts for future MG trials and clinical care.\u003c/p\u003e","manuscriptTitle":"Proteomic and Machine Learning Analysis Predicts Treatment Response Signatures in Myasthenia Gravis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-26 07:06:02","doi":"10.21203/rs.3.rs-7915342/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"","date":"2026-01-08T03:28:09+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-11-14T09:26:25+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-10-24T11:25:46+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of Translational Medicine","date":"2025-10-21T12:39:33+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"journal-of-translational-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jtrm","sideBox":"Learn more about [Journal of Translational Medicine](http://translational-medicine.biomedcentral.com)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/jtrm/default.aspx","title":"Journal of Translational Medicine","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"865c2eea-3003-4a21-97f8-00367702070f","owner":[],"postedDate":"November 26th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-03-22T18:03:01+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-26 07:06:02","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7915342","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7915342","identity":"rs-7915342","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}