Gut-brain nexus: Mapping multimodal links to neurodegeneration at biobank scale

Abstract

Alzheimer's disease (AD) and Parkinson's disease (PD) are influenced by genetic and environmental factors. We conducted a biobank-scale study to (i) identify endocrine, nutritional, metabolic, and digestive disorders with potential causal or temporal associations with AD/PD risk before diagnosis; (ii) assess plasma biomarkers' specificity for AD/PD in the context of co-occurring gut related traits and disorders; and (iii) integrate multimodal datasets to enhance AD/PD prediction. Our findings show that several disorders were associated with increased AD/PD risk before diagnosis, with variation in the strength and timing of associations across conditions. Polygenic risk scores reveal lower genetic predisposition for AD/PD in individuals with co-occurring disorders. Moreover, the proteomic profile of AD/PD cases was influenced by comorbid gut-brain axis disorders. Last, our multimodal prediction models outperform single-modality paradigms in disease classification. This endeavor illuminates the interplay between factors involved in the gut-brain axis and the development of AD/PD, opening avenues for therapeutic targeting and early diagnosis.

Fig. 1.. Study design.

The initial phase of our study used clinical data sourced from electronic medical records alongside genetic and proteomic data obtained from the UKB. Quality control procedures were rigorously applied to clinical and genetic datasets, including filtering for individuals of European ancestry, exclusion of related samples, and extraction of 155 ICD-10 codes representing diagnoses related to digestive, endocrine, nutritional, and metabolic disorders. Proteomic data underwent normalization of protein expression levels as part of quality control measures. The culmination of this phase involved the application of a Cox proportional hazards model, examination of polygenic risk scores (PRSs), and development of a generalized linear model (GLM). These analyses collectively contributed to the construction of a multimodal classification predictive model for AD and PD. Phase 2 of our study entailed validating these findings using data from the SAIL and FinnGen biobanks.

Fig. 2.. Receiver operating characteristic curve and balanced accuracy comparison for AD.

Performance evaluation of multiomics integration models using clinical, genetic, proteomic, and demographic data for AD.

Fig. 3.. Receiver operating characteristic curve and balanced accuracy comparison for PD.

Performance evaluation of multiomics integration models using clinical, genetic, proteomic, and demographic data for PD.

Fig. 4.. Feature importance plots.

(A) Distribution of the top 20 features that had the most substantial effect on the AD risk estimates. Each point represents a patient and the amount of effect on model output for each feature depends on its SHAP value. For example, the effect of the “Demographics/age_at_recruitment” feature on model output is large and positive (indicating a higher risk) when the patient has high values for “Demographics/age_at_recruitment” (more red points are on the right side). Similarly, (B) shows the top features for PD risk estimates. dCTP, 2′-deoxycytidine 5′-triphosphate.