Introducing myself

Introducing Our groups

Statistics & Bioinformatics and Nutrition & Metabolomics groups @ UB

Session objectives

• Overview of Patwhay Analysis for Metabolomics

• Introduce its components and

• Go through some methods with some detail
  

• Discuss some limitations and provide recomendations.

• Introduce some tools for Pathway Analysis

• Get a practical grasp of how to apply it.

Bioinformatics workflows

Metabolomics Worflow Example

From samples to features lists (2)

Metabolomics Workflows in MetaboAnalyst 5.0

An open problem: Metabolites IDs

• To be able to do Pathway Analysis, metabolites need to be mappable to their sources of information.

  • Must be uniquely identifiable by names/IDs.
  • Must be possible to link/relate these names/IDs with the corresponding IDs in the source of information we wish to rely.

• This is far from possible for all metabolites.

• Uniquely and unambiguosly naming all metabolites is, in the best of cases, “work in progress”.

Many sources for names

Some compund databases

Computed descriptors for Cholesterol

Computed descriptors for Cholesterol

What to do in practice (“real life”)

• Try to name your metabolites using standard ids from the very beginning

• Locate a dictionary or some id-converter that allows translating from one type of IDs to another when needed

• Keep in mind that many translations can end up in missing correspondence so decreasing your list size.

An example dictionary

And here comes Pathway Analysis

• Pathway Analysis (PWA) aims to interpret metabolomics results in terms of biological processes rather than individual compounds.
  • Instead of asking whether one metabolite changes,
  • It asks whether a coordinated group of metabolites associated with the same biochemical function shows evidence of perturbation.

The components of PWA

• With the previous goals in mind most PWA methods combines:
  • The list of features, with
  • Pre-existing sources of information related to them
• And, after some processing, it yields
  • some type of scores about
  • groups of features appearing to be coordinatedly related with the process being studied.

How can we interpret these lists?

From Lists to Biology

The elements of Pathways Analysis

• Loosely speaking, to do Pathway Analysis one needs:

  • A list of features, characterizing a process.

  • A source of information about these features.

  • An algorithm to highlight relevant information by linking list and source.

  • A tool implementing the algorithm.

• In this section, we focus on sources of information and on how to provide it to the algorithms.

Sources of information for PWA

Some common databases in Metabolomics

PWA and Metabolite Sets

Sources of Metabolite Sets

• Some sources of information directly provide metabolite sets. E.g: Chemical Ontologies, The KEGG Pathway Database

• For compound DBs, Metabolite sets may be built

  • By manual curation
  • Automatically, eg. using clustering approaches such as MetaMap or Chemical Similarity clustering

Chemical Ontologies

Metabolites Set libraries

Overview of MSEA’s metabolite set libraries

Metamap clusters

Chemical similarity clusters

The main idea of PWA

• Pathway Analysis (PWA) aims to interpret metabolomics results in terms of biological processes rather than individual compounds.

  • Instead of asking whether one metabolite changes,

  • It asks whether a coordinated group of metabolites associated with the same biochemical function shows evidence of perturbation.

• PWA will usually rely on (1) a list of interesting metabolites and (2) some type of knowledge source.

Types of Pathway Analysis

Khatri et alt. 10 years of Pathway Analysis

Targeted vs Untargeted

Comparison of targeted and untargeted metabolomics approaches highlighting key features. Videtič Paska, A. (2025) https://BioRender.com/7vhpmxv.

Obtaining feature lists

Example

Assess significance: Fisher test

• The example shows two cases
  • One where percentages are quite different
  • Another where percentages are similar.
• How can we set a threshold to decide that the difference is “big enough” to call it “Enriched”
  • Use Fisher Test or, equivalently,
  • a test to compare proportions or
  • a hypergeometric test.

Example 1: Surprisingly enriched

P-value small, odds-ratio high: List is surprisingly enriched in Feature Set

Example 2: Non-enriched

P-value high, odds-ratio around 1: List is not enriched in Feature Set

Posible problems with ORA

• No “natural” value for the threshold used to define significance
• Possible loss of statistical power due to thresholding
• No resolution between significant signals with different strengths
• Weak signals neglected
• Different results at different threshold settings.
• Based on the unrealistic assumption of independent feature (or feature group) sampling, which increases false positive predictions.

Enrichment Scores

GSEA may improve ORA

• A pathway can contain many metabolites changing moderately but in the same direction.

• ORA may fail here because very few metabolites cross the significance threshold.

• GSEA, can still detect these metabolites systematically concentrated near the top or bottom of ranking.

• Interpretation is more global than in ORA: coordinated patterns may be detected even if no single metabolite is significant by itself.

GSEA limitations

• The choice of ranking statistic matters, and different rankings can lead to different conclusions.
• The method can also be less intuitive than ORA because the enrichment score is not based on a simple overlap count.
• In metabolomics, where pathways are often small and annotation is incomplete, its performance may be less robust than in transcriptomics.

FELLA limitations

• The main limitation of FELLA is that it requires annotation, which means that it cannot be directly applied to raw untargeted peak lists.
• In addition, its results depend strongly on the quality, completeness and biological relevance of the network used. - Finally, topology-based methods are typically more computationally demanding and sometimes more difficult to explain than set-based methods (Khatri, Sirota, and Butte 2012; Lu, Pang, and Xia 2022).

(4): Mummichog

• What to do when you don’t know what the metabolites ions are?

• Most popular option is Mummichog (Li et al. 2013), a method that bypasses the requirement of full metabolite identification before pathway analysis.

Mummichog Method (1)

• The method starts from an untargeted peak list
• and builds pathway evidence before exact metabolite identification has been completed.

A mummichog pseudo-example

• Imagine a dataset with 1000 peaks where the top (most significant) 100 peaks are retained.

• Even if their identities are not known, many may still map to overlapping sets of candidate metabolites.

• If a relevant biological pathway is truly perturbed, those candidates will tend to accumulate in that pathway more strongly than expected by chance.

• The interpretation is therefore probabilistic and pathway-centered rather than metabolite-centered.

Mummichog: limitations

• Because multiple candidate metabolites can be assigned to each feature, the method may generate false positives.
• Its performance depends on the quality and coverage of the pathway database and on the quality of the mass matching step.
• The output is extremely useful for hypothesis generation, but it should not be confused with definitive metabolite identification.

Pathway Analysis Tools

A comparison of tools

The space of tools (in 2017)

Not the same, not that different

• ORA tools provided consistent results among tools revealing that these analyses are robust and reproducible regardless of their analytic approach.

• Redundancy of identifiers, Use of chemical class identifiers and Incompleteness of databases sets limit the extent of the analyses and reduce their accuracy.

• More work in the completeness of metabolite/pathway databases is required to get more accurate and global insights of the metabolome. # Summary, and all that

Summary

• Pathway Analysis is a useful approach to help gain biological understanding from omics-based studies.

• There are many ways, many methods, many tools

• Guide the choice by a combination of meaning, availability, ease of use and usefulness.

• Usually obtained from a good understanding of what it does and ow it is done.

• Different methods may yield different results.
  Worth checking!

Acknowledgements

Multiple testing

• Whatever approach we use for pathway Analysis there is a common characteristic: Every test is applied for every feature set in a long collection of sets

• This leads to a multiple testing problem: the Type I error probability of falsely rejecting the null hypothesis increases with the number of tests.

• In order to avoid an artificial inflation of False positive discoveries some adjustments are recommended.

Hypothesis Tests Decision Table

In a test with a null and an alternative hypothesis there are 2 possible right decisions and two possible incorrect ones (Type I and Type II errors)

Why Multiple testing matters

TYpe I error not useful here

How to deal with this issue?

Family Wise Error Rate

• Let \(M\) be the number of annotations tested.

• Given p-value, \(p\) compute \(p_{adj}=p\times M\), or

• Given significance level \(\alpha\) compute \(\alpha_{adj}=\alpha/M\).

• The adjusted P-value, \(p_{adj}\) is greater than or equal to the probability that one or more of the observed enrichments are due to random draws.

• This adjustment is said to controling for the Family-Wise Error Rate (FWER).

• Bonferroni method controls FWER.

Bonferroni Caveats

• This adjustment is very stringent and can “wash away” real enrichments leading to false negatives,

• Often one is willing to accept a less stringent condition, that is accepting some false positives to avoid too many false negatives.

• This is may be done using the “false discovery rate” (FDR), which leads to a gentler correction when there are real enrichments.

False Discovery Rate

• FDR is the expected proportion of “False Positives” that is of the observed enrichments due to chance.
• Less restrictive than Bonferroni adjustment which is a bound on the probability that any one of the observed enrichments could be due to random chance.
• Typically, FDR adjustments are calculated using the Benjamini-Hochberg procedure.
• FDR threshold is often called the “q-value”