Skip to content
/ Metabolomics-in-Alzheimer-s-Disease---An-Investigation-into-Conflicting-Methodologies-and-Results Public

Metabolomics Data Science MSc project looking at significant metabolites found in Alzheimer's Disease (AD) patients in comparison to patients with mild cognative imparement (MCI) and cognative normal (CN) controls.

2 stars 0 forks Branches Tags Activity
Star
Notifications You must be signed in to change notification settings

VikkiWalls/Metabolomics-in-Alzheimer-s-Disease---An-Investigation-into-Conflicting-Methodologies-and-Results

BranchesTags

Repository files navigation

Metabolomics in Alzheimer's Disease - An Investigation into Conflicting Methodologies and Results

Data Science MSc project within the Metabolomics research area.

The project looks at significant metabolites found in Alzheimer's Disease (AD) patients in comparison to patients with mild cognitive impairment (MCI) and cognitive normal (CN) controls. It also compares results to existing literature which already lacks a consensus on which metabolites are significant to this disease.

Table of Contents

  1. General Info
  2. Technologies
  3. Methods
  4. Results
  5. List of Documents
  6. References

General Info

Metabolomics is a field of research which is new and fast growing. It looks at biological compounds extracted from minimally invasive tissue and fluid samples such as blood plasma, urine and cerebral spinal fluid (CSF). These compounds can be used to determine a person’s health status as well as what they are at risk of.

The field is very data heavy and its newness makes it the perfect environment for innovation. It’s an ideal space for data science projects with a focus on social/public good.

This project takes an existing dataset consisting of 1909 metabolites and three groups (CN, MCI and AD) of 15 individuals, produced by the Mayo Clinic in 2013 and uses various statistical and machine learning methods to extract metabolites significant to Alzheimer’s disease (AD).

In existing literature there is little in the way of overlap in results from study to study, and thus whilst AD has been covered a number of times in the short life of the field of Metabolomics, there is no defined list of metabolites which are significant to this disease and backed up over multiple studies. In this paper, the significant metabolites were compared against metabolites found significant by other literature with the hope of providing some clarity to the noise.

Aims

This study aims to find metabolites with a statistically significant influence on individuals with AD from the existing dataset and to compare these results with the metabolites named as significant in other studies. This is done in the hope to either provide clarity by aligning with existing results or to illustrate that a lack of agreed "good practice" techniques within the metabolomics field results in inconsistent and irreproducible results.

Challenges

One of the biggest challenges was related to the specific dimensions of the dataset, with far more metabolites than observations (individual samples). This caused interference in initial attempts at statistical analysis and is the suspected cause for the unusual results in the VIF test. A related issue was the minimal numbers of sample sizes. With only 15 individual samples per cognitive type there is a reasonable possibility that any metabolites found to be significant may not occur in other datasets - this seems to be a common issue with the metabolomics field in general, and there certainly seems to be a lack of consensus in significant metabolites for AD throughout the existing literature.

Other common issues in this field relate to a lack of consensus around best practices in data analysis methods. This is addressed by Lee and Styczynski (2018) and was the given motivation behind the development of the No Skip k Nearest Neighbour (NS-kNN) method of imputation. This in itself brought its own challenge as the code produced by Lee and Styczynski was in the Matlab Language. Wilst efforts were made both to translate from Matlab into Python, and to export the data from Python into Matlab and back to run the function, none of these efforts worked and so this method, whilst clearly the better option than kNN, had to be disincluded from the final project.

According to the Human Metabolome Database there are a total of 253,245 metabolites in the human body, of this 3,444 have been detected and quantified and 20,924 have been detected but not quantified. This leaves a massive proportion of metabolites which may potentially be significant to AD untested. The dataset used in these analyses only contained 1,909 metabolites. Until it is possible to test all 253,245 metabolites for any metabolomics study (i.e., including wider than AD research) there will always be a risk that results are not wholly relevant.

Technologies

A list of technologies used within the project:

Whilst Matlab was not used in the final project, attempts to use the NS-kNN method (prior to it's disinclusion) used the Matlab Version R2022a.

Methods

Statistical and Machine Learning Methods

  • Variance Inflation Factor (VIF) Testing

    • Used in conjunction with additional visualisation techniques mentioned below to test for multicollinearity

  • Principal Component Analysis (PCA)

    • Unsupervised Machine Learning Method

    • Finds the optimum number of components to explain the maximum variance in the data

    • Useful for high dimensions of features (e.g., 1909 metabolites)

    • Useful for simplifying complex data when there are high levels of multicollinearity

      A new dataframe was created using the principal components to explain 95% of the variance, and this dataframe was used in the PLS-DA test and stored as a csv file.

  • Partial Least Squares - Discriminant Analysis (PLS-DA)

    • Supervised Machine Learning Method
    • Used to determine whether known groups are actually different
    • Used to determine which features best describe the differences between the known groups
    Drawbacks

  • Logistic Regression

    • Supervised Machine Learning method (borrowed from statistics)
    • One of the most common methods of analysis in metabolomics
    • Useful for prediction and classification
    • Builds a regression model to predict the probability that a given data entry belongs to the category numbered as “1”

    Cross Validation (CV) was then used to assess accuracy and performance of the model.

    Feature selection by using the feature importances property of the sklearn.ensemble extra trees classifier was then used to build a second logistic regression model, which was again accuracy tested using CV.

Visualisation Techniques

  • Multicollinearity

    • Cluster Map (using Seaborn)
      • Can be difficult to interpret with large datasets such as metabolomics data
    • Interactive Network Graph (using NetworkX)
      • Adjustable correlation threshold (recommended to be set at 0.95) allows direct comparisons between different levels
      • Panning, zoom and selection tools make interpretation and manipulation accessible
      • Visualisation of direct relationships between correlated metabolites
      Drawbacks
      • Threshold slider must be moved once after restarting kernel to produce a display
      • Can still be difficult to interpret if high levels of multicollinearity exist and plot is zoomed out (making it difficult to use in academic write ups)
      • Large metabolite names can create visual noise

  • PCA

    • Scatter Graph (using Yellowbrick) - colour coded by cognitive status
    • Elbow Visualiser (using matplotlib) - explained variance against principal components

  • PLS-DA

    • Visual comparison between AD and CN, MCI and CN and, AD and MCI
    • Initial line plot using principal components to compare potential for outliers
    • Scatter plot based on PLS regression scores to check for separation between the profiles of each of the three groups

  • Logistic Regression

    • Bar chart showing the feature importance score of the top 50 variables (metabolites) for each model (AD vs CN and MCI vs CN)

Notes

One method was attempted however was not included in the final project due to lack of functioning code:

  • No Skip k Nearest Neighbour (NS-kNN)

    • Used for imputation in dealing with missing values
    • Specifically designed to improve accuracy for values in data missing not at random (MNAR)
    • Particularly useful for metabolite data where missing values can be due to Limit of Detection (LOD) levels in mass spectrometry (i.e., these values are MNAR)

Results

Multicollinearity

  • Visualisation

    • Clustermap displayed a clear clusters of highly positively correlated metabolites
      • Individual metabolites could not be identified due to large size of dataset
    • Interactive network graph was able to visualise direct relationships between correlated metabolites
      • e.g., the direct relationship between C5H6N4O4S and AsnProLys

  • VIF Test

    • Displayed infinity values for every metabolite, regardless of the specific programming method used
      • Normal techniques to reduce/remove this error (including log transforming the data, and removing the highest correlated values) did not change these results
    • Did not appear to match up with the correlation values produced by the correlation matrix used by the visualisation methods above

Whilst the VIF test appears to be inconclusive, between the VIF results and the visualisation techniques, it was clear that there were some levels of multicollinearity in this dataset. In order to undertake analysis methods such as logistic regression, this needed to be removed.

PCA

  • Visualisation

    • Production of a scatter graph colour coded by cognitive status revealed that data contained no specific clusters
    • Elbow plot did not show a clear change to indicate a specifically optimal number of principle components
  • Explained Variance Ratio indicated that the first 40 (out of a potential 45) components would explain 95% of the variance in the data
    • The first 29 components would explain 80% of the variance

PLS-DA

  • Alzheimer's Disease against Cognitive Normal

    • Initial visual comparison of components by cognitive type suggested a lack of obvious outliers although there were a couple of clear differences in spike height (but these did not go out with the range of the rest of the data)
    • On plotting the PLS Regression scores there is visible separation between AD and CN values on Latent Variable 1

  • Mild Cognitive Impairment against Cognitive Normal

    • Initial visual comparison of components by cognitive type indicated some potential outliers in the CN data
    • On plotting the PLS Regression scores there is visible separation between MCI and CN values on Latent Variable 1

  • Alzheimer's Disease against Mild Cognitive Impairment

    • Initial visual comparison of components by cognitive type indicated some potential outliers in the AD data
    • On plotting the PLS Regression scores there is clear and significant visible separation between AD and MCI values on Latent Variable 1

Logistic Regression

List of Documents

  • Document 1 - README.md (this file)
  • Document 2 - Full project code
  • Document 3 - Additional tables
  • Document 4 - Data

References

Methodological Sources

Boccard, J. and Rutledge, D.N. (2013). A consensus orthogonal partial least squares discriminant analysis (OPLS-DA) strategy for multiblock Omics data fusion. Analytica Chimica Acta, [online] 769, pp.30–39. doi:10.1016/j.aca.2013.01.022.

Brereton, R.G. and Lloyd, G.R. (2014). Partial least squares discriminant analysis: taking the magic away. Journal of Chemometrics, [online] 28(4), pp.213–225. doi:10.1002/cem.2609.

“Database Statistics: Metabolite Statistics.” Human Metabolome Database, The Metabolomics Innovation Centre, hmdb.ca/statistics. Accessed 3 Aug. 2022. HMD Release 5.0 - January 2022.

Lee and Styczynski, M.P. (2018). NS-kNN: a modified k-nearest neighbors approach for imputing metabolomics data. Metabolomics, [online] 14(12), p.153. doi:10.1007/s11306-018-1451-8.

Lee, M. and Hu, T. (2019). Computational Methods for the Discovery of Metabolic Markers of Complex Traits. Metabolites, 9(4), p.66. doi:10.3390/metabo9040066.

Ruiz-Perez, D., Guan, H., Madhivanan, P., Mathee, K. and Narasimhan, G. (2020). So you think you can PLS-DA? BMC Bioinformatics, [online] 21(S1). doi:10.1186/s12859-019-3310-7.

Schneider, Lon S., and Mary Sano. “Current Alzheimer’s Disease Clinical Trials: Methods and Placebo Outcomes.” Alzheimer’s & Dementia, vol. 5, no. 5, Sept. 2009, pp. 388–397, 10.1016/j.jalz.2009.07.038. Accessed 23 May 2022.

Szymańska, E., Saccenti, E., Smilde, A.K. and Westerhuis, J.A. (2011). Double-check: validation of diagnostic statistics for PLS-DA models in metabolomics studies. Metabolomics, [online] 8(S1), pp.3–16. doi:10.1007/s11306-011-0330-3.

Literature

Cao, B., Wang, D., Pan, Z., Brietzke, E., McIntyre, R.S., Musial, N., Mansur, R.B., Subramanieapillai, M., Zeng, J., Huang, N. and Wang, J. (2019). Characterizing acyl-carnitine biosignatures for schizophrenia: a longitudinal pre- and post-treatment study. Translational Psychiatry, 9(1). doi:10.1038/s41398-018-0353-x.

Clish, C.B. (2015). Metabolomics: an emerging but powerful tool for precision medicine. Cold Spring Harbor Molecular Case Studies, [online] 1(1). doi:10.1101/mcs.a000588.

Cummings, J. (2017). Lessons Learned from Alzheimer Disease: Clinical Trials with Negative Outcomes. Clinical and Translational Science, [online] 11(2), pp.147–152. doi:10.1111/cts.12491.

Douaud, G., Groves, A.R., Tamnes, C.K., Westlye, L.T., Duff, E.P., Engvig, A., Walhovd, K.B., James, A., Gass, A., Monsch, A.U., Matthews, P.M., Fjell, A.M., Smith, S.M. and Johansen-Berg, H. (2014). A common brain network links development, aging, and vulnerability to disease. Proceedings of the National Academy of Sciences, 111(49), pp.17648–17653. doi:10.1073/pnas.1410378111.

He, Y., Yu, Z., Giegling, I., Xie, L., Hartmann, A.M., Prehn, C., Adamski, J., Kahn, R., Li, Y., Illig, T., Wang-Sattler, R. and Rujescu, D. (2012). Schizophrenia shows a unique metabolomics signature in plasma. Translational Psychiatry, [online] 2(8), pp.e149–e149. doi:10.1038/tp.2012.76.

Human Metabolome Database. (2021). Human Metabolome Database: Showing metabocard for Hypoxanthine (HMDB0000157). [online],[Accessed 28 Jun. 2022].

Huo, Z., Yu, L., Yang, J., Zhu, Y., Bennett, D.A. and Zhao, J. (2020). Brain and blood metabolome for Alzheimer’s dementia: findings from a targeted metabolomics analysis. Neurobiology of Aging, [online] 86, pp.123–133. doi:10.1016/j.neurobiolaging.2019.10.014.

Illig, T., Gieger, C., Zhai, G., Römisch-Margl, W., Wang-Sattler, R., Prehn, C., Altmaier, E., Kastenmüller, G., Kato, B.S., Mewes, H.-W., Meitinger, T., de Angelis, M.H., Kronenberg, F., Soranzo, N., Wichmann, H.-E., Spector, T.D., Adamski, J. and Suhre, K. (2010). A genome-wide perspective of genetic variation in human metabolism. Nature Genetics, [online] 42(2), pp.137–141. doi:10.1038/ng.507 163 metabolites in plasma.

Kaddurah-Daouk, R. (2006). Metabolic Profiling of Patients with Schizophrenia. PLoS Medicine, [online] 3(8), p.e363. doi:10.1371/journal.pmed.0030363.

Manchester, M. and Anand, A. (2017). Chapter Two - Metabolomics: Strategies to Define the Role of Metabolism in Virus Infection and Pathogenesis. Advances in Virus Research, [online] 98, pp.57–81.

Mittelstrass, K., Ried, J.S., Yu, Z., Krumsiek, J., Gieger, C., Prehn, C., Roemisch-Margl, W., Polonikov, A., Peters, A., Theis, F.J., Meitinger, T., Kronenberg, F., Weidinger, S., Wichmann, H.E., Suhre, K., Wang-Sattler, R., Adamski, J. and Illig, T. (2011). Discovery of Sexual Dimorphisms in Metabolic and Genetic Biomarkers. PLoS Genetics, [online] 7(8), p.e1002215. doi:10.1371/journal.pgen.1002215 Metabolite inclusion criteria for ‘Schizophrenia shows a unique metabolomics signature in plasma’.

Oliver, Stephen G., et al. “Systematic Functional Analysis of the Yeast Genome.” Trends in Biotechnology, vol. 16, no. 9, 1 Sept. 1998, pp. 373–378. Trends in Biotechnology, 10.1016/s0167-7799(98)01214-1. Accessed 3 Aug. 2022.

Petersen, R.C. (2004). Mild cognitive impairment as a diagnostic entity. Journal of Internal Medicine, [online] 256(3), pp.183–94. doi:10.1111/j.1365-2796.2004.01388.x.

Proitsi, P., Kim, M., Whiley, L., Simmons, A., Sattlecker, M., Velayudhan, L., Lupton, M.K., Soininen, H., Kloszewska, I., Mecocci, P., Tsolaki, M., Vellas, B., Lovestone, S., Powell, J.F., Dobson, R.J.B. and Legido-Quigley, C. (2016). Association of blood lipids with Alzheimer’s disease: A comprehensive lipidomics analysis. Alzheimer’s & Dementia, [online] 13(2), pp.140–151. doi:http://dx.doi.org/10.1016/j.jalz.2016.08.003.

Ribe, A.R., Laursen, T.M., Charles, M., Katon, W., Fenger-Grøn, M., Davydow, D.S., Chwastiak, L., Cerimele, J.M. and Vestergaard, M. (2015). Long-term Risk of Dementia in Persons With Schizophrenia: A Danish Population-Based Cohort Study. JAMA Psychiatry, [online] 72(11), pp.1–7. doi:10.1001/jamapsychiatry.2015.1546.

Rubin, E. (2016). The Relationship Between Schizophrenia and Dementia | Psychology Today. [online] www.psychologytoday.com. [Accessed 24 May 2022].

Trushina, E., Dutta, T., Persson, X.-M.T., Mielke, M.M. and Petersen, R.C. (2013). Identification of Altered Metabolic Pathways in Plasma and CSF in Mild Cognitive Impairment and Alzheimer’s Disease Using Metabolomics. PLoS ONE, [online] 8(5), p.e63644. doi:10.1371/journal.pone.0063644.

Wang, G., Zhou, Y., Huang, F.-J., Tang, H.-D., Xu, X.-H., Liu, J.-J., Wang, Y., Deng, Y.-L., Ren, R.-J., Xu, W., Ma, J.-F., Zhang, Y.-N., Zhao, A.-H., Chen, S.-D. and Jia, W. (2014). Plasma Metabolite Profiles of Alzheimer’s Disease and Mild Cognitive Impairment. Journal of Proteome Research, [online] 13(5), pp.2649–2658. doi:10.1021/pr5000895.

White, K.E. and Cummings, J.L. (1996). Schizophrenia and Alzheimer’s disease: Clinical and pathophysiologic analogies. Comprehensive Psychiatry, 37(3), pp.188–195. doi:10.1016/s0010-440x(96)90035-8.

Yu, Z., Kastenmüller, G., He, Y., Belcredi, P., Möller, G., Prehn, C., Mendes, J., Wahl, S., Roemisch-Margl, W., Ceglarek, U., Polonikov, A., Dahmen, N., Prokisch, H., Xie, L., Li, Y., Wichmann, H.-E., Peters, A., Kronenberg, F., Suhre, K. and Adamski, J. (2011). Differences between Human Plasma and Serum Metabolite Profiles. PLoS ONE, [online] 6(7), p.e21230. doi:10.1371/journal.pone.0021230 163 Metabolites in Plasma (NB: Some authors the same as Schizophrenia Data source. Potentially questionable credibility given authors referenced themselves).

About

Metabolomics Data Science MSc project looking at significant metabolites found in Alzheimer's Disease (AD) patients in comparison to patients with mild cognative imparement (MCI) and cognative normal (CN) controls.

Topics

python metabolomics alzheimers-disease ns-knn

Resources

Readme
Activity

Stars

2 stars

Watchers

1 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

No packages published

Languages