Pathway-based machine learning analysis of Parkinson’s disease transcriptomics data reveals coordinated alterations in inflammatory pathways

Aug 4, 2026

Neuroinflammation is implicated in Parkinson’s disease progression, yet the specific molecular pathways remain unclear. This study applies statistical and machine learning analyses to cross-sectional and longitudinal transcriptomics from PD patients and controls, comparing gene-level changes with aggregated functional representations at pathway, cellular compartment, and protein-complex levels. Differential expression and time-correlation analyses address confounders, while nested cross-validation and SHAP identify predictive inflammatory features.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages1

Reading Time3 min

AccessFree to download

Background

Methods

Data representations and feature aggregation

Differential expression and time correlation

Machine learning evaluation and interpretability

Results

Time-related inflammation pathway correlations

Performance of aggregated features

SHAP-identified predictive complexes and pathways

Conclusions

Pathway-based machine learning analysis of Parkinson’s disease transcriptomics data reveals coordinated alterations in inflammatory pathways ELISA GÓMEZ DE LOPE, ENRICO GLAAB ON BEHALF OF THE NCER-PD CONSORTIUM [elisa.gomezdelope@uni.lu](elisa.gomezdelope@uni.lu) Biomedical Data Science Group, LCSB, University of Luxembourg @elisagdelope Background Neuroinflammation has been implicated in the progression of Parkinson's disease (PD) by contributing to dopaminergic neuron loss 1 , but the specific molecular pathways involved remain largely unknown. We applied statistical and machine learning (ML) analyses to cross-sectional and longitudinal transcriptomics data from PD patients and controls, examining both gene level changes and aggregated functional representations, such as pathway-, cell compartment-and protein complex-level features. | | Methods | | | | | --- | --- | --- | --- | --- | | | | | | | |
Higher level functional (pathway, cellular location and protein complex) representations of transcriptomic data from the PPMI cohort2 (whole blood) were generated using aggregation statistics (mean, median, sd) and low-dimensional representations using PCA and principal curves3 (Pathifier software).
Differential expression analyses accounting for the confounders age and gender, and time correlation analyses (A) were applied to PD case/control data for both gene and functional level representations. Next, we evaluated ML models using a nested cross-validation including the feature selection and parameter optimization for PD versus control sample classification, assessing the performance of pathway-level aggregation statistics and single-level features . A SHAP value analysis was conducted to identify the most informative predictive features . | | | Abundance matrix (mxn)

GOBP, GOCC, CORUM Database Mappings | Aggregation based on database mappings
Higher aggregated features
Aggregated abundance matrix (m x p)

Aggregation statistics, PCA & Principal curves | level • Inflammation pathways displayed PD-specific positive correlation with time (A) . Results • Features reflecting the variance of aggregated expression at protein complex level (CORUM) provided higher cross-validated performance (AUC) than other types of aggregations and in comparison to the original gene features for PD vs . control sample classification . Logistic regression provided an area under the curve (AUC) of 0.67±0.06 in this setting (B) . • In these ML models, multiple pathways, cellular locations and protein complexes involved in (neuro)inflammation were included among the most relevant predictive features (C) . | • A SHAP value analysis for the logistic regression model applied to CORUM-sd aggregated features revealed (neuro)inflammation, mitochondrial and chromatin modification complexes among the feature sets with the highest relevance . Furthermore, the SMN protein complex involved in the survival of motor neurons and neuron degeneration was included among the most predictive feature sets (D) . | | | | --- | --- | --- | | A. Longitudinal mean-aggregated counts at GO BP level

Negative regulation of humoral immune response
Complement receptor mediated signaling pathway
T0 T 1 T2 T3
Mean-aggregated counts Mean-aggregated counts | C. Aggregated functional representations associated with inflammation and neuro-inflammation among the top-20 most relevant features for corresponding ML models according to a SHAP value analysis of feature relevance .
GO BP (Pathway level) GOCC (Cellular compartment) CORUM (Protein complex) MEAN Innate immune response in mucosa
Negative regulation of intrinsic apoptotic signaling in response to DNA damage Chitosome
Postsynaptic endosome Platelet dense tubular network
Platelet alpha granule membrane Endocytic vesicle lumen TNF-alpha/NF-kappa B signaling complex (CHUK, KPNA3, NFKB2, NFKBIB, REL, IKBKG, NFKB1, NFKBIE, RELB, NFKBIA, RELA, TNIP2) HRD1

FAQs

What does the study investigate about Parkinson’s disease transcriptomics?

It examines coordinated molecular alterations linked to neuroinflammation by analyzing transcriptomics across multiple feature levels, including pathways, cellular compartments, and protein complexes.

How were transcriptomic features represented for machine learning?

Higher-level functional representations were generated by aggregating statistics (mean, median, sd) and using low-dimensional representations such as PCA and principal curves via Pathifier.

Which features were most predictive for PD vs controls?

Protein-complex-level aggregated variance (CORUM) achieved higher cross-validated performance than other aggregation types and single gene features, with logistic regression reaching an AUC of 0.67±0.06.

Aug 4, 2026

Upload by Lucas Martin