1. Computational Challenges in Metabolomics (Part 1)
David Wishart, University of Alberta
Dagstuhl Seminar on Computational Mass Spectrometry
Schloss Dagstuhl, Germany Aug , 2015

2. The Pyramid of Life Genome Metabolomics Proteomics Genomics Proteome
Metabolome
Physiological Influence
Environmental Influence
Proteome
Genome

3. Why Small Molecules Count
100% of all agricultural products (herbicides, pesticides, fertilizers) are small molecules
>99% of all compounds that give food or drinks their aroma, color and taste are small molecules
91% of all known drugs are small molecules
>85% of all common clinical assays test for small molecules
60% of all drugs are derived from pre-existing metabolites
10-15% of identified genetic disorders involve diseases of small molecule metabolism

4. Proteomics vs. Metabolomics

5. Proteomics vs. Metabolomics
Very MS or MS/MS oriented
Good separation is critical
Generates lots of raw data
Peptide and protein ID
Isotopic labeling (ICAT) helps
Possible to derive 3D structure
Permits protein imaging
Very dependent on databases
Spectral processing and deconvolution is challenging
Quantitation is challenging
Data analysis requires MV stats
Data integration is challenging
Better software is needed
Very MS or MS/MS oriented
Good separation is critical
Generates lots of raw data
Chemical ID
Isotopic labeling (SIL) helps
Possible to derive 3D structure
Permits metabolite imaging
Very dependent on databases
Spectral processing and deconvolution is challenging
Quantitation is challenging
Data analysis requires MV stats
Data integration is challenging
Better software is needed

6. Proteomics vs. Metabolomics

7. Proteomics Workflow Biofluid/Extracts HPLC or PAGE Tryptic Digest
MALDI plate
Protein ID
Mass Fingerprint
MS analysis

8. Protein ID by PMF-MS

9. Metabolomics Workflow
Biological or Tissue Samples
Extraction
Biofluids or Extracts
Compound ID
LC/GC-MS Spectra
LC-MS or GC-MS

10. Compound ID by GC/LC-MS
LC/GC-MS total
Ion chromatogram
CH3

11. Proteomics vs. Metabolomics
Polymers of 20 amino acids (chemically similar)
185 million sequences (from DNA sequencing)
Sequence defines MS & MS/MS spectra
Trypsin gives definable cleavages
MS alone can ID proteins (PMF)
MS/MS fragmentation at 1 fixed energy
MS/MS fragmentation is easily predictable and very distinct
30 common PTMs
PTMs are somewhat predictable
1000s of distinct chemical classes (chemically diverse)
No information from DNA sequencing
Structure defines MS & MS/MS spectra (adducts, fragments)
No trypsin for small molecules (CID only)
MS alone cannot ID metabolites
Different energies for different molecules
MS/MS & EI-MS fragments not easily predictable, often similar
>400 PTMs via metabolism
PTMs are hard to predict

12. Challenges for Metabolomics
Most MS-based metabolomics studies ID <100 cmpds (<1% of the known metabolome)
Metabolite ID requires accurate, referential MS/MS or EI-MS spectra and/or RT information
Limited experimental MS/MS, EI-MS & RT data
The chemical space of most metabolomes is not fully known (perhaps >5 million compounds total)
<1% of the chemicals in PubChem are relevant to metabolomics
Metabolomics needs specialized compound and spectral (MS/MS, EI-MS, NMR) databases
Metabolomics needs computational tools to predict biologically viable metabolites and their spectra

13. LC-MS Spectral DBs MoNA – 236,604 spectra, 69,946 cmpds** (12,000)
METLIN – 68,124 spectra, 13,048 cmpds
mzCloud – 422,349 spectra, 2975 cmpds
NIST14 MS/MS – 234,284 spectra, 9344 cmpds
MassBank – 28,185 spectra, 11,500 cmpds
Wiley LC-MSn – >10,000 spectra, 4500 poisons
ReSpect – 9107 spectra, 3595 cmpds
GNPS – 9000 spectra, 4200 natural products
Total #compounds with exp. MS/MS spectra ~20,000
Less than 60% are biologically relevant

14. How to Get Missing Spectra?
Obtain or synthesize all biologically relevant molecules (metabolites, HPVs, drugs, pollutants, foods, etc.), prepare or synthesize their metabolites and collect their NMR, LC-MS and GC-MS spectra COST - 5,000,000 cmpds X $1000/cmpd = $5 billion OR
Do this entire exercise computationally COST - 5,000,000 cmpds X $0.10/cmpd = $500,000

15. Computational Metabolomics
Predicted biotransformations
(50,000 --> 5,000,000)
Known biomolecules (50,000)
Match observed spectra
to predicted spectra to ID compounds
Predicted MS/MS, NMR, GC-MS
Spectra of knowns + biotransformed

16. The Human Metabolome Database (HMDB)
A web-accessible resource containing detailed information on 41,993 “quantified”, “detected” and “expected” metabolites
Data mined from the literature and other eDBs
100’s of drug metabolites
1000’s of xenobiotics
>10,000 reference spectra
Supports sequence, spectral, structure and text searches as well as compound browsing
Full data downloads

17. The Drug Database (DrugBank v. 4.3)
1602 small molecule drugs
>5000 experimental drugs
Data mined from the literature and other eDBs
>1000 drugs with metabolizing enzyme data
>1200 drug metabolites
>600 MS+NMR spectra
>4200 unique drug targets
208 data fields/drug
Supports sequence, spectral, structure and text searches as well as compound browsing
Full data downloads

18. The Toxic Exposome Database (T3DB)
Comprehensive data on toxic compounds (drugs, pesticides, herbicides, endocrine disruptors, drugs, solvents, carcinogens, etc.)
Data mined from the literature and other eDBs
>3600 toxic compounds
>1900 reference spectra
~2100 toxic targets
Supports sequence, spectral, structure, text searches as well as compound browsing
Full data downloads

19. Computational Metabolomics
Predicted biotransformations
(50,000 --> 5,000,000)
Known biomolecules (50,000)
Match observed spectra
to predicted spectra to ID compounds
Predicted MS/MS, NMR, GC-MS
Spectra of knowns + biotransformed

20. Secondary Metabolism Diazepam Tempazepam Oxazepam Nordazepam
CH3
Tempazepam
Oxazepam
Nordazepam
Diazepam
N-(2-Benzoyl-4-chlorophenyl)-2-acetamidoacetamide

21. BioTransformer

22. BioTransformer - Flowchart
Query Molecule
Other Reactions
Phase I
Reaction-specific structural constraints
Enzyme metabolite?
(Machine Learning)
YES
YES
YES NO
SOM Predictor
(Machine Learning)
Metabolite Generator NO
SOMs NO
Metabolites
All structures are generated as
SMILES, SDF or MOL files
No metabolites

23. BioTransformer – SOM Prediction
Preference Learning based on 100 atomic (e.g. atom type) and 10 molecular features (e.g. mass)
SOM predictor was trained for 9 CYP450s
Average Prediction accuracy of 84.54%
Structures generated based on 92 Phase I reactions

24. BioTransformer Results?
6,230 Phase I metabolites ?
9,510 Phase II metabolites
5,000 compounds ?
6,110 Microbial metabolites ?
12,340 Other metabolites
34,000 metabolites
~220,000

25. Computational Metabolomics
Predicted biotransformations
(50,000 --> 5,000,000)
Known biomolecules (50,000)
Match observed spectra
to predicted spectra to ID compounds
Predicted MS/MS, NMR, GC-MS
Spectra of knowns + biotransformed

26. Computational Challenges in Metabolomics (Part 2)
Sebastian Böcker, Friedrich Schiller University
Dagstuhl Seminar on Computational Mass Spectrometry
Schloss Dagstuhl, Germany Aug , 2015