Protein-based Diagnosis and Analysis of Co-pathologies Across Neurodegenerative Diseases: Large-Scale AI-Boosted CSF and Plasma Classification

Abstract

Neurodegenerative diseases (including Alzheimer's disease, Parkinson's disease, Frontotemporal dementia, and Dementia with Lewy bodies) pose diagnostic challenges due to overlapping pathology and clinical heterogeneity. We leveraged proteomic data from more than 21,000 cerebrospinal fluid and plasma samples to develop and validate explainable, boosting-based multi-disease AI classifiers. The models achieved weighted AUCs in the testing datasets of 0.97 for CSF and 0.88 for plasma, equivalent to traditional biomarkers. The model was validated with neuropathological and clinical data, confirming robust generalizability without any retraining. Using zero-shot learning, we classified disease subtypes including autosomal dominant AD and prodromal PD and clarified disease states for those with conflicting clinical information. The model also showed the ability to prioritize cognitively normal individuals at disease risk. This framework enabled the identification and quantification of continuous, individual-level disease probabilities that allow for the quantification of overlap across diseases and co-pathologies within an individual. Through this work, we establish a benchmark computational framework for enhancing diagnostic precision in NDs.

Figure 1.. Study Design. A. Sample Demographics.

CSF samples from the training & testing datasets were obtained from the Knight ADRC, ADNI, Stanford ADRC, MDC, Barcelona-1, Fundacio ACE, and PPMI. Plasma samples from the training and testing datasets were obtained from the Knight ADRC, Stanford ADRC, and MDC. Samples used for validation were accessed through DIAN, ROSMAP, the I-ADRC, and the GNPC. Samples used for reclassification were from the Knight-ADRC & ROSMAP. B. Workflow Steps. Proteomic data (SOMAscan 5k or 7k) from five main diagnostic categories, two tissues, and multiple cohorts (Knight-ADRC, ADNI, Stanford ADRC, Barcelona-1, Fundacio ACE, MDC, and PPMI) was used to identify differentially abundant proteins in each disease compared to controls. Those proteins were then filtered based on their importance and were used to build a tree-based gradient boosting AI disease classifier (LightGBM) using a trainingtesting paradigm. Using independent external data from ROSMAP and the I-ADRC along with related disease categories from internal data, validation of the model performance was done using relevant cognitive and neuropathological phenotypes. Finally, the classifier was used to clarify diagnoses for controls with biomarker evidence of disease and individuals with unclear disease etiology and to investigate specific samples with characteristics of interest, Control; AD, Alzheimer’s Disease; PD, Parkinson’s Disease; FTD, Frontotemporal Dementia; DLB, Dementia with Lewy Bodies, ADAD, autosomal dominant Alzheimer’s disease; PDD: Prodromal PD; OT, other dementia; MCI, mild cognitive impairment.

Figure 2.. Differential Protein Expression in CSF and Plasma Across Neurodegenerative Diseases.

(A, B) Heat Maps of Averaged Protein Expression in CSF and Plasma. Average z-scored expression profiles of 400 cerebrospinal fluid (CSF) analytes (A) and 700 plasma analytes (B) used for downstream classification, spanning AD, PD, FTD and DLB. Rows represent individual analytes, while columns represent mean expression values for each disease category. Warmer colors (red) indicate higher relative expression levels, and cooler colors (blue) indicate lower relative expression. (C, D) Disease-Specific Analyte Heat Maps (Top 15 per Disease). In C (CSF) and D (Plasma), the top 15 nominally significant proteins for each disease are arranged by effect size, yielding a combined total of 60 analytes. Color intensity reflects the absolute value of the effect size, with red indicating higher expression or stronger positive associations and blue indicating lower expression or negative associations. Asterisks highlight significant changes (* p < 0.05; ** FDR < 0.05), underscoring potential biomarker candidates for differential diagnosis. (E, F) UpSet plots of F-test–selected analytes. Panels E (CSF) and F (plasma) display the intersections among the final feature sets obtained after variance filtering and F-test ranking. The CSF panel summarizes 400 retained analytes, whereas the plasma panel summarizes 700 analytes. Bars above each plot indicate the number of analytes shared by the corresponding disease combinations, highlighting both disorder specific markers and analytes common to multiple neurodegenerative conditions.

Figure 3.. CSF and Plasma Multi-panel Performance.

(A) CSF Classifier Performance. The upper flow chat shows the schematic overview of the cerebrospinal fluid (CSF) workflow, beginning with differential abundance analysis and subsequent F-test-based feature selection to narrow the set to 400 top-ranked analytes. A LightGBM model with SMOTETomek oversampling was then trained for multi-class discrimination among CO, AD, PD, FTD, and DLB. The lower panels show CSF internal 30% hold-out testing performance. Panel 1: clinical status for samples included in the CSF testing dataset. Panel 2: Radar plot depicting the probability distributions (mean and 95% CI) for each of the five major status categories. Panel 3: AUC curves for classification accuracy for each disease alongside the model’s overall weighted AUC (0.97) and macro AUC (0.95). Panel 4: Whisker plot for AUCs for all five classes with 95% confidence interval (CI). Panel 5: Heatmap summarizes both AUC-ROC and AUC-PR for the five classes. (B) Plasma Classifier Performance. Again, the upper flow chart showed the schematic overview of the plasma workflow, initiated with differential abundance analysis (3,607 disease-associated analytes) followed by variance- and F-test-based feature selection to yield 700 top-ranked analytes. A LightGBM model with SMOTETomek oversampling was then trained to classify CO, AD, PD, FTD, and DLB. The lower multi-panels showed plasma internal 30% hold-out testing performance. Panel 1: clinical status for samples included in the plasma testing dataset. Panel 2: Radar plot depicts probability distributions (mean and 95% CI) for each of the five major status categories. Panel 3: AUC curves for classification accuracy for each disease alongside the model’s overall weighted AUC of 0.88 (macro AUC 0.79). Panel 4: Whisker plot for AUCs for all five classes with 95% CI. Panel 5: Heatmap again shows AUC-ROC and AUC-PR for the five classes.

Figure 4.. External validation and clinical pathological correlates of the proteomic classifier.

(A) CSF zero-shot validation performance. Radar plots depict mean class probability profiles (AD, PD, DLB, FTD, CO) for internal CSF samples not used in training or testing: autosomal dominant AD (ADAD, n = 114), prodromal PD (n = 332) and Parkinson’s disease dementia (PDD, n = 13). (B) CSF clinical and pathological correlations. Predicted AD probability stratified by: Clinical Dementia Rating (CDR) at baseline within the DIAN cohort, final CDR within the internal dataset, Braak neurofibrillary tangle (NFT) stage and neuritic plaque C score. Bars show mean ± s.e.m; higher clinical or pathological burden aligns with higher AD probability. (C) Plasma cohort level performance. Radar plots for categories from ROSMAP or held-out diagnoses from internal data. ROSMAP: cognitively unimpaired controls (n = 507), neuropathology confirmed AD (n = 167) and PD (n = 2). Knight ADRC: ADAD carriers (n = 125) and PDD (n = 92). (D) Concordance between model assignments and ROSMAP neuropathology findings. Left panels: linear fits (95 % CI) of AD probability versus neuritic plaque burden and neurofibrillary tangle burden across brain regions (higher is worse AD pathology). Right panels: mean probabilities (± s.e.m) across Amyloid and Braak stage, which depend on plaque and tangle abundance. (E) Concordance between model assignments and ROSMAP clinical findings. Scatter plots with regression lines relate AD probability to Mini Mental State Examination (MMSE) and episodic memory zscores. Bar charts show mean probabilities (± s.e.m.) across clinical cognitive diagnosis strata and clinical PD status.

Figure 5.. AI-based reclassification of ambiguous clinical diagnosis.

(A) CSF clinical controls. Left: Radar plot showing mean class-probability profiles (AD, PD, DLB, FTD, CO) for participants clinically labeled as CO at enrollment, separated by CSF AD biomarker status (amyloid and tau, AT). Right: Bar charts display the distribution of CSF AT status among samples the model retained as CO versus those reclassified as AD. (B) Plasma clinical controls. Radar plot and bar charts analogous to panel A, using plasma-based class probabilities and pTau217 positivity (T+/T–). (C) CSF “Other” diagnoses (OT). Same analyses as in panel A, applied to individuals with indeterminate clinical labels. The radar plot reveals mixed probability profiles, while bar charts compare AT status between samples reclassified as CO versus AD. (D) Plasma “Other” diagnoses. Radar plot and bar charts stratified by pTau217 status illustrate concordance between AI-based classification and biological markers. (E) ROSMAP mild cognitive impairment (MCI). Panel 1: Radar plot of class probabilities for MCI participants. Panels 2 & 3: AD probability plotted against amyloid PET Centiloid scores and post-mortem Braak NFT stages. Panel 4: Mean class probabilities (± s.e.m.) stratified by clinical Parkinson’s disease certainty levels. (F) Transition plots. Samples from the GNPC are shown with their clinical diagnoses (left) and their AI-based classifications (right) in CSF and plasma.