Gut Microbiome Changes in Alzheimer's Disease: A Multi-Model Machine Learning Analysis

Script:

Hey guys, my name is Jason Zhang. And my name is Elsa Gibson. We’re grade 12 students at Mount Douglas Secondary School.

Our project focuses on a new way to diagnose Alzheimer’s. Alzhimers is incredibly common, affecting nearly 2% of the entire population. We chose it because it is also very difficult to diagnose often requiring invasive and expensive tests.

But fear not, a fecal test could be on the horizon! The bacteria living in your gut, the microbiome, can directly influence your brain, and vice versa, kind of like a “gut feeling”. Our project used a machine learning model to detect alzhimers using data from the bacteria in your gut which is collected by a…. Fecal sample! (a poo sample)

“Wanna hear what we found? Come check out our project either online or at our poster at the Canada Wide National Science in Edmonton!”

From forgetting once hard-etched names of loved ones to forgetting how to tell time, Alzheimer’s Disease is more than just the seventh leading cause of death globally; it is a disabling disease that mainly affects the older population. AD makes up 60-70% of almost 800,000 dementia cases in Canada, which will only continue to rise given Canada’s rapidly aging population. AD is a neurodegenerative disease characterized by neuronal damage caused by the formation of neurofilamentary tangles (NFTs) composed of Tau protein, and β-amyloid plaques that accumulate and block synaptic connections. However, AD diagnosis can be an arduous process that often involves physical neurological examination, followed by numerous tests to rule out other diseases, such as blood tests or brain scans, which are expensive and often have lengthy waiting lists. Cerebrospinal fluid (CSF) tests detecting Tau or β-amyloid are used, but can be invasive.

Various neurological disorders, including AD, have been linked to changes in the gut microbiome, the community of bacteria living inside the colon. The brain is connected to the gut via the microbiota-gut-brain-axis, a bidirectional transportation system of endocrine, neural, and metabolic signals which are affected by the bacteria. Microbial imbalance, called dysbiosis, has been associated with poorer blood-brain barrier permeability, increased β-amyloid aggregation, and increased neuroinflammation, which could contribute to AD. Additionally, AD has been linked to lower bacterial diversity in the microbiome, as well as abnormal abundances of certain bacteria.

Could AD be accurately diagnosed using data taken from a fecal sample?

Fecal metagenomic sequencing of 75 Alzheimer’s disease (AD) and 100 healthy controls (HC) patients was obtained from the European Nucleotide Archive (Accession Code: PRJEB47976). The sequencing reads were profiled using databases to quantify bacterial genera, gene ontology (GO), and KEGG Orthology (KO) pathways. Lastly, a CLR transformation was applied to account for the mathematical constraints of microbiome compositional data.

After processing, the analysis was split into three blocks:

Block A: Diversity Analysis

Diversity factors in the number of different metabolites or taxa and their relative abundance. 2 types of diversity were analyzed. Alpha Diversity is the diversity of a sample’s gut microbiome. It was analyzed via Simpson and Shannon scores. Beta Diversity is the difference in diversity between samples. It was analyzed via PERMANOVA & PCoA.

Block B: Significance Comparison

Microbiome association analysis was done using the MaAsLin2 R package. This statistical method is used to estimate the significance of genera, KO, & GO features in AD, accounting for metadata. MaAsLin2 ranked all features to identify the most AD-associated, using patients' sex, age, and APOE status as metadata.

Block C: Machine Learning (ML)

The goal for the ML was to diagnose AD as accurately as possible. RandomForest, ElasticNet, and SVM were trained on increasing amounts of KO, GO and genus data, starting with the most significant from the association analysis. A 5-fold cross-validation (CV) splits the data into 5 equal parts. Then all models are trained 5 times, each time using a different part as its validation set, and the remaining 4 as its training (See slide 3). Inside each fold, an association analysis was performed to determine the top features. This novel method reduces data leakage, improving this study over previous work. Lastly, performance was compared to the patient's APOE gene, the current genetic predictor for AD.

Block A: Ecological Analysis

Analysis of GO, KO, and genus features showed that Alpha diversity between AD and HC patients had no significant difference (p = 0.8576, p = 0.6696) based on Shannon and Simpson scores evaluated by Wilcoxon. This implies that overall gut diversity is preserved in AD, which is consistent with other studies, including the Wisconsin Microbiome in Alzheimer's Risk Study. (Kang et al. 2025)

Beta diversity highlighted differences between groups for GO and KO. Genus PERMANOVA found no statistically significant difference (p = 0.669). KO and GO PERMANOVAs were statistically significant (p = 0.039, p = 0.038, respectively).

This showed that taxonomic composition does not significantly differentiate AD from HC, but functional composition does. Recent studies have shown that while taxonomic composition is altered in dementia, metabolic signals are much stronger (Zhao et al., 2025). Additionally, the AlzBiom study found that using KO signals was superior to taxonomic data for detecting AD (Laske et al. 2025). Taxonomic composition can vary with diet and geographic location, while KO reflects the gut metabolic environment, providing a more stable and consistent signal for disease detection.

Block B: Significance comparison

KO and GO features were generally more significant. From the association analysis, 1111 of 5357 (20.7%) KO features had a p-value under 0.05, compared to 914 GO features of 4677 (19.5%), and 8 genera of 509 (1.5%) (See slide 3).

Of the KO and GO features, some were higher in HC and others in AD. This supports the neuroinflammatory ideas discussed by Seo & Holtzman (2024). The dysbiosis of AD involves the enrichment of pro-inflammatory pathways, not just the depletion of protective ones.

Based on the top 20 KO features, the most prominent changes occurred in the anaerobic respiration pathway and redox-balancing proteins. Reduction of the Hydrogenase-4 complex (K12138, K12143, K12136)–responsible for preventing oxidative stress from acidic gut environments–increased the oxidative stress and the accumulation of reactive oxygen species (ROS). Although these changes happen in the gut, the leaky gut theory suggests that the ROS can drain into the bloodstream and lymphatic system. When ROS enters the brain, it increases inflammation, which plays a role in AD.

Block C: Machine Learning

The models performed well, with AUC scores around or higher than most of the current literature. The Random Forest model with 8 features emerged as the most optimal configuration, achieving a mean AUC of 0.774 ± 0.09 with a high recall of 0.798, which is significant for clinical screening. Comparatively, the AlzBiom study had AUC scores of ~0.78, but with more biomarkers. Vogt. et al. 2017 only obtained an AUC of 0.72, but their work was novel and did not use KO features.

Crucially, model performance was independent of a patient's APOE status. The performance remained the same in APOE4-Negative subgroups. These results suggest that AD-associated dysbiosis is not merely a bystander, but part of a bidirectional feedback loop. Neurological changes influence shifts in the gut microbiome, which then accelerate neurodegeneration.

A fecal test could detect AD with a 79.8% recall score. Despite the small dataset used, these results are a proof of concept for a new detection tool in predicting AD. If a large enough dataset were collected from people across the world, a model could theoretically be built to predict AD clinically.

A theoretical detection tool would follow the subsequent steps:

Collect a fecal sample at a Patient Services Centre (e.g. LifeLabs).

Send the sample to a DNA sequencing laboratory to extract, prepare, and read the DNA into base pairs.

Compare the DNA sequence to known databases to convert to features for ML.

Run a RandomForest model trained on data similar to the patient.

There are many benefits to using a fecal test before other diagnostic procedures.

Cost

A facility such as BC Children’s Hospital’s “Gut for Health” lab could perform steps 2 and 3 for less than $100. This price point rivals other testing methods; MRI, CT, or PET scans can cost upwards of $1000, as can CSF tests.

Accessibility

Patient Services Centres are commonplace across Canada. These offer shorter wait times and walk-in appointments. This is a massive improvement over other diagnostic tests, which can be booked out for months.

Early Detection

If AD is caught early, there are options for disease-modifying therapies to be used to slow the progression of symptoms. This test may provide a method to detect AD at an earlier stage of cognitive impairment.

The main issue with our project is that it is purely correlative. If we had access to a longitudinal dataset, we could measure changes in the gut microbiome over time and relate them to the progression of AD. This could allow us to build a model for early detection and disease progression monitoring. Additionally, this would aid in finding possible causative taxa or metabolites, which could be studied further for therapeutics. We have reached out to the Canadian Longitudinal Study on Aging for access to their data, which could provide a large, longitudinal dataset with a focus on Canadian patients.

Our project would never have happened without the tremendous support of those around us. Thank you to parents, Joel Gibson and Gina Capretta and to our teachers, Dave Newell, Neal Johnson, Amelita Kucher, and Phil Ferriera, for supporting us throughout by reading over our project and pushing us to constantly improve. Thank you to the VIRSF sponsors for supporting our project and allowing us to attend CWSF free of charge. Lastly, thank you for taking the time to engage with our project and support young scientists!

Dataset:

Laske C, Müller S, Preische O, Ruschil V, Munk MHJ, Honold I, Peter S, Schoppmeier U, Willmann M. Signature of Alzheimer’s disease in intestinal microbiome: Results from the ALZBiOM study. Frontiers in Neuroscience. 2022;16:792996.

Reference Materials:

1. Laske C, Müller S, Preische O, Ruschil V, Munk MHJ, Honold I, Peter S, Schoppmeier U, Willmann M. Signature of Alzheimer’s disease in intestinal microbiome: Results from the ALZBiOM study. Frontiers in Neuroscience. 2022;16:792996. https://doi.org/10.3389/fnins.2022.792996. doi:10.3389/fnins.2022.792996

2. Vogt NM, Kerby RL, Dill-McFarland KA, Harding SJ, Merluzzi AP, Johnson SC, Carlsson CM, Asthana S, Zetterberg H, Blennow K, et al. Gut microbiome alterations in Alzheimer’s disease. Scientific Reports. 2017;7(1):13537. https://doi.org/10.1038/s41598-017-13601-y. doi:10.1038/s41598-017-13601-y

3. National Institute on Aging. Beyond the brain: The gut microbiome and Alzheimer’s disease. National Institute on Aging. 2023 Jun 12. https://www.nia.nih.gov/news/beyond-brain-gut-microbiome-and-alzheimers-disease

4. What is Alzheimer’s? Alzheimer’s Association. https://www.alz.org/alzheimers-dementia/what-is-alzheimers

5. Seo D-O, Holtzman DM. Current understanding of the Alzheimer’s disease-associated microbiome and therapeutic strategies. Experimental & Molecular Medicine. 2024;56(1):86–94. https://www.nature.com/articles/s12276-023-01146-2. doi:10.1038/s12276-023-01146-2

6. Mayo Clinic Staff. Alzheimer’s Disease - Diagnosis & treatment. Mayo Clinic. 2026 Mar 3. https://www.mayoclinic.org/diseases-conditions/alzheimers-disease/diagnosis-treatment/drc-20350453

7. National Institute on Aging. What happens to the brain in Alzheimer’s disease? National Institute on Aging. 2026 Jan 27. https://www.nia.nih.gov/health/alzheimers-causes-and-risk-factors/what-happens-brain-alzheimers-disease

8. Heredia L, Mateo D, Carrión N, Torrente M. Dataset on neuropsychological profile and microbiota composition in cognitively unimpaired elderly and Alzheimer’s patients. Data in Brief. 2025;61:111778. https://doi.org/10.1016/j.dib.2025.111778. doi:10.1016/j.dib.2025.111778

9. Murray ER, Kemp M, Nguyen TT. The Microbiota–Gut–Brain Axis in Alzheimer’s Disease: A Review of taxonomic Alterations and Potential Avenues for Interventions. Archives of Clinical Neuropsychology. 2022;37(3):595–607. https://doi.org/10.1093/arclin/acac008. doi:10.1093/arclin/acac008

10. Professional CCM. The Gut-Brain connection. Cleveland Clinic. 2026 Apr 14. https://my.clevelandclinic.org/health/body/the-gut-brain-connection

11. Liu S, Gao J, Zhu M, Liu K, Zhang H-L. Gut microbiota and dysbiosis in Alzheimer’s disease: Implications for pathogenesis and treatment. Molecular Neurobiology. 2020;57(12):5026–5043. https://doi.org/10.1007/s12035-020-02073-3. doi:10.1007/s12035-020-02073-3

12. Li Z-L, Ma H-T, Wang M, Qian Y-H. Research trend of microbiota-gut-brain axis in Alzheimer’s disease based on CiteSpace (2012–2021): A bibliometrics analysis of 608 articles. Frontiers in Aging Neuroscience. 2022;14:1036120. https://doi.org/10.3389/fnagi.2022.1036120. doi:10.3389/fnagi.2022.1036120

13. Liang Y, Liu C, Cheng M, Geng L, Li J, Du W, Song M, Chen N, Yeleen TAN, Song L, et al. The link between gut microbiome and Alzheimer’s disease: From the perspective of new revised criteria for diagnosis and staging of Alzheimer’s disease. Alzheimer S & Dementia. 2024;20(8):5771–5788. https://doi.org/10.1002/alz.14057. doi:10.1002/alz.14057

14. Kang JW, Khatib LA, Heston MB, Dilmore AH, Labus JS, Deming Y, Schimmel L, Blach C, McDonald D, Gonzalez A, et al. Gut microbiome compositional and functional features associate with Alzheimer’s disease pathology. Alzheimer S & Dementia. 2025;21(7):e70417. https://doi.org/10.1002/alz.70417. doi:10.1002/alz.70417

15. Raulin A-C, Doss SV, Trottier ZA, Ikezu TC, Bu G, Liu C-C. ApoE in Alzheimer’s disease: pathophysiology and therapeutic strategies. Molecular Neurodegeneration. 2022;17(1):72. https://doi.org/10.1186/s13024-022-00574-4. doi:10.1186/s13024-022-00574-4

16. KAAS - KEGG Automatic Annotation Server. https://www.genome.jp/kegg/kaas/

17. Gene Ontology overview. Gene Ontology Resource. 2026 Apr 23. https://geneontology.org/docs/ontology-documentation/

18. Services – Gut4Health. https://www.bcchr.ca/gut4health/services/

Images:

19. Research Highlights on Health and Aging. Statistics Canada. 2016 Jul 28. https://www150.statcan.gc.ca/n1/pub/11-631-x/11-631-x2016001-eng.htm

20. Mayo Clinic Staff. Alzheimer’s Disease - Symptoms & Causes. Mayo Clinic. 2026 Mar 3. https://www.mayoclinic.org/diseases-conditions/alzheimers-disease/symptoms-causes/syc-20350447

21. Megur A, Baltriukienė D, Bukelskienė V, Burokas A. The Microbiota–Gut–Brain axis and Alzheimer’s Disease: neuroinflammation is to blame? Nutrients. 2020;13(1):37. https://doi.org/10.3390/nu13010037. doi:10.3390/nu13010037

22. humann2 – The Huttenhower Lab. https://huttenhower.sph.harvard.edu/humann2/

23. KEGG PATHWAY: Alzheimer disease - Hom0 sapiens (human). https://www.kegg.jp/pathway/hsa05010

24. Embl-Ebi. What is GO? | GOA and QuickGO. https://www.ebi.ac.uk/training/online/courses/goa-and-quickgo-quick-tour/what-is-go/

25. Pelletier H. How-To: Cross Validation with Time Series Data. Towards Data Science. 2025 Jan 22. https://towardsdatascience.com/how-to-cross-validation-with-time-series-data-9802a06272c6/

26. Stool Collection Tubes with DNA Stabilizer. https://www.invitek.com/en/microbiome/pdp-stool-collection-tubes-with-dna-stabilizer

27. Canadian Longitudinal Study on Aging (CLSA). Research Platform for Health & Aging - Canadian Longitudinal Study on Aging (CLSA). Canadian Longitudinal Study on Aging (CLSA). 2026 Jan 5. https://www.clsa-elcv.ca/

Models:

28. maaslin2 – The Huttenhower Lab. https://huttenhower.sph.harvard.edu/maaslin2/

29. Breiman L. Random forests. Machine Learning. 2001;45(1):5–32. https://doi.org/10.1023/a:1010933404324. doi:10.1023/a:1010933404324

30. Cortes C, Vapnik V. Support-vector networks. Machine Learning. 1995;20(3):273–297. https://doi.org/10.1007/bf00994018. doi:10.1007/bf00994018

31. Zou H, Hastie T. Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society Series B (Statistical Methodology). 2005;67(2):301–320. https://doi.org/10.1111/j.1467-9868.2005.00503.x. doi:10.1111/j.1467-9868.2005.00503.x