Data Science and Environmental Cheminformatics (SanDAL Workshop, Uni Lu)

(1)

1

Assoc. Prof. Dr. Emma L. Schymanski

FNR ATTRACT Fellow and PI in Environmental Cheminformatics

Luxembourg Centre for Systems Biomedicine (LCSB), University of Luxembourg

Email: emma.schymanski@uni.lu and @ESchymanski

Plus ECI-LSCB, NORMAN, PubChem, IPB, Eawag, UFZ and many other

colleagues who contributed to our science over the years!

Talk available under DOI: 10.5281/zenodo.4075013

SanDAL Data Science Workshop, October 13-14, 2020, Belval Campus environment environmental

sample

high resolution

mass spectrometry cheminformatics identify chemicals and toxic effects

Data Science and

Environmental Cheminformatics

(2)

2

Source: https://en.wikipedia.org/wiki/File:EU-Luxembourg.svg

.eu

(3)

3

Example: NORMAN Collaborative Trial 2015

Schymanski et al. 2015, ABC, DOI: 10.1007/s00216-015-8681-7

Croatian

Water

RWS

(4)

4

Example: ELIXIR (https://elixir-europe.org/)

Schymanski et al. 2015, ABC, DOI: 10.1007/s00216-015-8681-7

(5)

5

Environmental Cheminformatics

@

Luxembourg Centre for

Systems Biomedicine

LCSB-ECI, June 2019/2020

(6)

6

Environmental Cheminformatics

@

Luxembourg Centre for

Systems Biomedicine

LCSB-ECI, June 2019 + extras

(7)

7

1 10 100 1000 10000 100000 1 million 1 billion chemicals …. …. ….

Our (Community) Challenge: Identifying Chemicals

Schymanski et al. (2014) DOI: 10.1021/es4044374; Vermeulen et al. (2020) DOI: 10.1126/science.aay3164

Sample

High resolution

mass spectrometry

(8)

8

Our (Community) Challenge: Identifying Chemicals

Sample

High resolution

mass spectrometry

Chemicals

AND connecting

chemical knowledge

(9)

9

Our Context (“worst case”): The Exposome

Image c/o uni.lu (LCSB); modified from Vermeulen et al (2020) Science. DOI: 10.1126/science.aay3164

At its most complete, the exposome

encompasses life-course environmental

exposures (including lifestyle factors),

from the prenatal period onwards.

Christopher P. Wild, 2005.

DOI: 10.1158/1055-9965.EPI-05-0456

(10)

10

Our Aim: Generating Insight from Measurements

Mod. From Fig. 2, Sevin et al 2015, Current Opinion in Biotechnology. DOI: 10.1016/j.copbio.2014.10.001

Analytical methods

Automated spectral data

processing

Data interpretation

Hypothesis generation

(11)

11

General Scheme(s) of Mass Spectrometry

Top: chemwiki.ucdavis.edu. Bottom H. Röst & M. Steiner. Own work. https://commons.wikimedia.org/w/index.php?curid=11640370

Information about “parent” Information about “parent”

and fragments

(substructures)

(12)

12

Mass Spectra are our “fingerprints”

https://metabolomics-usi.ucsd.edu/spectrum/?usi=mzspec:MASSBANK::accession:EQ300803

(13)

13

But … this is what the output “really” looks like …

Image © www.planetorbitrap.com/q-exactive

(14)

14

From Sample to Numbers (Masses) to Molecules

o Identification = turning numbers into structures

N

N N

S CH₃

NH NH

CH₃

C H₃

CH₃

N N N

S CH₃

NH NH

C H₃

CH₃ O H

P O S

S O

C H₃

CH₃ P OH S

S

O CH₃

CH₃

OH

C H₃

S O O OH

CH₃

C H₃

S

N S O

O OH

S O

O OH

CH₃ CH₃

S O

O OH

C H₃

CH₃

S O

O OH C

H₃

C H₃

S O

O OH C

H₃

CH₃

S O

O OH C

H₃

CH₃ N N N

S

NH NH

CH₃ C H₃

CH₃

NH₂ OH O

massbank.eu

(15)

15

How to measure? LC vs GC?

Brack et al 2016. DOI: 10.1016/j.scitotenv.2015.11.102

(16)

16

How to measure? Ionisation

Singh, RR et al (2020) DOI: 10.1007/s00216-020-02716-3; Ulrich, EM et al (2019) DOI: 10.1007/s00216-018-1435-6

n=1264

(17)

17

How to measure? Widely Varying Concentrations

Rappaport et al (2014) Environmental Health Perspectives. DOI: 10.1289/ehp.1308015

(18)

18

Our challenge? We have many unknowns …

(l) Data from Schymanski et al 2014, ES&T DOI: 10.1021/es4044374. (r) E. coli data provided by N. Zamboni, IMSB, ETH Zürich.

W aste w ate r Ce lls

(19)

19

…and insufficient reference data (mass spectra)

H. Oberacher et al. (2020) Environmental Sciences Europe 32: 43. DOI: 10.1186/s12302-020-00314-9

o Mass spectra are our “fingerprints” for identification

o Only available for ~0.1-4 % of Exposomics-relevant resources

(20)

20

What we measure (in dust): LC vs GC

Rostkowski et al 2019. DOI: 10.1007/s00216-019-01615-6

LC-MS

(n=969)

GC-MS

(n=592)

(21)

21

What we measure in biota/water/sediment

Alygizakis NA et al (2019) DOI: 10.1016/j.trac.2019.04.008 Interactive heatmap at http://norman-data.eu/NORMAN-REACH

Retrospective screening of REACH chemicals in

Black Sea samples (various matrices)

(22)

22

Emerging contaminants in different countries

Alygizakis NA et al (2018) DOI: 10.1021/acs.est.8b00365and DOI: 10.5281/zenodo.2623815

(23)

23

NORMAN Digital Sample Freezing Platform

“Live” retrospective screening of known and unknown

chemicals in European samples (various matrices)

http://norman-data.eu/ and Alygizakis NA et al (2019) DOI: 10.1016/j.trac.2019.04.008.

(24)

24

Real-time Monitoring of the Rhine River

Hollender, Schymanski, Singer & Ferguson, 2018, ES&T Feature, 51:20, 11505-11512. DOI: 10.1021/acs.est.7b02184

Previously unknown chemicals detected due to “stand-out” patterns

(25)

25

Historical Contamination in Lake Sediments

Chiaia-Hernandez et al. Anal. Bioanal. Chem. 2014, 406 (28), pp 7323-7335. DOI: 10.1007/s00216-014-8166-0

(26)

26

Micropollutant Time Trends in

Albergamo et al. 2019 Environ. Sci. Technol. 53 (13), 7584-7594, DOI: 10.1021/acs.est.9b01750

Riverbank Filtration Systems

(27)

27

Analytical Methods: Target vs Non-target

KNOWNS SUSPECTS No Prior Knowledge

HPLC separation and HR-MS/MS

TARGET

ANALYSIS

SUSPECT

SCREENING

NON-TARGET

SCREENING

Targets found Suspects found Masses of interest

(Molecular formula)

DATABASE

SEARCH

STRUCTURE

GENERATION

Confirmation and quantification of compounds present

Candidate selection (retention time, MS/MS, calculated properties)

Sampling extraction (SPE) HPLC separation HR-MS/MS

Time, Effort & Number of Compounds….

SUSPECTS

SPECTRUM

SEARCH

Spectral match

(28)

28

Our Aim: Generating Insight from Measurements

Analytical methods

Automated spectral data

processing

Data interpretation

Hypothesis generation

(29)

29

Target List Suspect List

(e.g. NORMAN,

LMC, Eawag-PPS,

ReSOLUTION)

Componentization

(nontarget)

TARGET

ANALYSIS

SUSPECT

SCREENING

NON-TARGET

SCREENING

(enviMass,

vendor software)

Gather evidence

(nontarget,

ReSOLUTION,

RMassBank)

Masses of interest

Molecular formula

determination

(enviPat, GenForm)

Non-target identification

(MetFrag2.3, ReSOLUTION)

Sampling extraction (SPE) HPLC separation HR-MS/MS

Detection of blank/blind/noise/internal standards; time trend analysis (enviMass)

Conversion (Proteowizard) and Peak Picking (enviPick, xcms, MZmine, …)

Prioritization

(enviMass)

MS/MS Extraction

(RMassBank)

Interpretation, confirmation, peak inventory, confidence and reporting

Altenburger et al, 2019, Env. Sci. Europe. DOI: 10.1186/s12302-019-0193-1

(30)

30

Non-target Identification: Where to start?

Componentization

(nontarget)

NON-TARGET

SCREENING

Masses of interest

Molecular formula

determination

(enviPat, GenForm)

Non-target identification

(MetFrag)

Prioritization

(enviMass)

MS/MS Extraction

(RMassBank)

Directed by the

scientific question

(31)

31

Targets, Non-targets and Isotopes (ESI-) by Intensity

Schymanski et al. 2014, ES&T, 48: 1811-1818. DOI: 10.1021/es4044374

Artificial Sweeteners

Diclofenac

Pictures: www.coca-cola-com; www.rivella.ch; www.voltargengel.com

(32)

32

Non-target Identification: Where to start?

Componentization

(nontarget)

NON-TARGET

SCREENING

Masses of interest

Molecular formula

determination

(enviPat, GenForm)

Non-target identification

(MetFrag)

Prioritization

(enviMass)

MS/MS Extraction

(RMassBank)

Gather

Experimental

Evidence

(33)

33

Grouping Adducts and Isotopes

0 3000

6000

9000

12000

15000

positive

2%

27%

100%

Noise/Blank Targets Non-targets

0 3000

6000

9000

12000

15000

positive

negative

1%

30%

100%

(34)

34

Gathering Evidence for Identification I

Determining the molecular mass and elements, adducts present

388.2551

[M+NH

₄

]

⁺

Schymanski et al. 2015, ABC,

DOI: 10.1007/s00216-015-8681-7

M. Loos, et al. 2015, DOI:

10.1021/acs.analchem.5b00941

http://www.eawag.ch/forschung/uchem/software/

http://cran.r-project.org/web/packages/nontarget/

http://cran.r-project.org/web/packages/enviPat/

(35)

35

Gathering Evidence for Identification II

Detecting the presence of elements with enviPat and nontarget

13 C

15 N

13 C, ¹⁸ O

34 S

M. Loos, et al. 2015, DOI: 10.1021/acs.analchem.5b00941

http://www.envipat.eawag.ch/ http://cran.r-project.org/web/packages/enviPat/

Image © www.seanoakley.com/

(36)

36

Detecting Isotope Signals in Samples

“Classic” environmental strategy: Cl isotopes

A small number of

¹⁵

N isotopes

A surprising number of

³⁴

S isotopes

Cl

N

S

(37)

37

Peak Grouping in Workflows

This can be done automatically with nontarget / enviMass (and others)

(38)

38

Non-target Identification: Where to start?

Componentization

(nontarget)

NON-TARGET

SCREENING

Masses of interest

Molecular formula

determination

(enviPat, GenForm)

Non-target identification

(MetFrag)

Prioritization

(enviMass)

MS/MS Extraction

(RMassBank)

Gather

Evidence

(39)

39

Targets, Non-targets and Isotopes (ESI-)

S O

O O^-

m/z = 79.96

(40)

40

MS2: Extracting Mass Spectra I

Automatic MS and MS/MS

Recalibration and Clean-up

Remove interfering peaks

Spectral Annotation with

- Experimental Details

- Compound Information

https://github.com/MassBank/RMassBank/

http://bioconductor.org/packages/RMassBank/

Stravs, Schymanski, Singer and Hollender, 2013,

Journal of Mass Spectrometry, 48, 89–99. DOI: 10.1002/jms.3131

(41)

41

MS2: Extracting Mass Spectra II

Anjana Elapavalore, Mira Narayanan,

Todor Kondic, Jessy Krier,,

Hiba Mohammed Taha.

https://git-r3lab.uni.lu/eci/shinyscreen

(42)

42

Non-target Identification: Next Step: Candidates!

Componentization

(nontarget)

NON-TARGET

SCREENING

Masses of interest

Molecular formula

determination

(enviPat, GenForm)

Non-target identification

(MetFrag)

Prioritization

(enviMass)

MS/MS Extraction

(RMassBank)

MS1:

213.9637 1168637.0 [M-H]^- 214.9630 14790.0 M+1/15N 214.9670 94077.2 M+1/13C 215.9595 104643.6 M+2/34S 215.9679 7060.2 M+2/13C

MS/MS:

134.0054 339689.4 150.0001 77271.2 213.9607 632466.8

(43)

43

Our Aim: Generating Insight from Measurements

Analytical methods

Automated spectral data

processing

Data interpretation

Hypothesis generation

(44)

44

Our (Community) Challenge: Identifying Chemicals

Sample

High resolution

mass spectrometry

Chemicals

AND connecting

chemical knowledge

(45)

45

Our Toolkit: NORMAN-SLE: Capturing Expert Knowledge

https://www.norman-network.com/nds/SLE/

Large, merged / market lists

Intermediate lists / specialised collections

> 73 lists

> 144,980 substances

Medium lists Small, highly

specialized lists

(46)

46

Our Toolkit: NORMAN-SLE: Capturing Expert Knowledge

https://www.norman-network.com/nds/SLE/

Large, merged / market lists

Intermediate lists / specialised collections

PFAS

Pesticides

Pharma

Priority Lists

CCS

Smoke

Dust

Water

Plastics

Surfactants

REACH

Toxins

NPS

> 73 lists

> 144,980 substances

(47)

47

Our Toolkit: MassBank EU – Mass Spectra

http://massbank.eu/MassBank und https://github.com/MassBank/

>88,100 spectra

~16,500 compounds

>47 contributors

(48)

48

Our Toolkit: PubChem – Finding the “known unknowns”

Kim S, Chen J, Cheng T, et al. PubChem 2019 update: …. Nucleic Acids Res. 2019;47(D1):D1102–D1109. doi:10.1093/nar/gky1033

https://pubchem.ncbi.nlm.nih.gov/

(49)

49

Our Toolkit: MetFrag – Annotating/Identifying Masses

Ruttkies, Schymanski, Wolf, Hollender, Neumann (2016) J. Cheminf., 2016, DOI: 10.1186/s13321-016-0115-9

(50)

50

MetFrag + PubChem + MS/MS only

5 ppm

0.001 Da

mz [M-H]

^-

213.9637

± 5 ppm

MS/MS

134.0054 339689

150.0001 77271

213.9607 632466

Status: 2010

Ranked Candidates

(51)

51

Challenge: MetFrag + PubChem + MS/MS only

Ruttkies et al. (2016) DOI: 10.1186/s13321-016-0115-9, Bolton & Schymanski (2020), DOI: 10.5281/zenodo.3611238& in prep.

o MS/MS doesn’t provide sufficient information alone …

MetFragRL, PubChem 2016 MS/MS only (n=473)

0 20 40 60 80 100%

(52)

52

MS/MS does not provide sufficient information alone…

MetFrag + PubChem + Formula Search + https://massbank.eu/MassBank/RecordDisplay.jsp?id=EQ300804&dsn=Eawag

(53)

53

MetFrag – MS/MS and MORE!

5 ppm

0.001 Da

mz [M-H]

^-

213.9637 or

PubChem

± 5 ppm

RT: 4.54 min

355 InChI/RTs

References

Tox. Data

Data Sources

Exposure Info

MS-ready links

Suspect Lists

MS/MS

134.0054 339689

150.0001 77271

213.9607 632466

Elements: C,N,S

S O O

OH

Status: 2016

(54)

54

MetFrag + PubChem + MS/MS + Metadata

o Adding literature, references & RT boosts to ~71 % rank 1!

MetFragRL, PubChem 2016 MS/MS only (n=473) MetFragRL, PubChem 2016 MS/MS + Metadata (n=1298)

0 20 40 60 80 100%

(55)

55

Connecting multiple lines of evidence for identification

MetFrag + PubChem + Formula + MoNA + SusDat + Pat + Refs + https://massbank.eu/MassBank/RecordDisplay.jsp?id=EQ300804&dsn=Eawag

(56)

56

Connecting multiple lines of evidence for identification

MetFrag + PubChem + Formula + MoNA + SusDat + Pat + Refs + https://massbank.eu/MassBank/RecordDisplay.jsp?id=EQ300804&dsn=Eawag

Candidates with high information content

Candidates with low information content

Challenge: the growing number of candidates …

Need: wide coverage and high efficiency!

(57)

57

Exposomics Use Case:

PubChemLite tier1: 360 K

The 111 million Challenge …

Bolton & Schymanski (2020). PubChemLite tier0 and tier1 (Version PubChemLite.0.2.0) DOI: 10.5281/zenodo.3611238

Environmental Use Case:

PubChemLite tier0: 316 K

111 million … OR …

the most relevant / annotated?

(58)

58

PubChemLite: tailor-made database + metadata

MetFrag+PubChemLite+Formula+MoNA+SusDat+Pat+Refs+Anno + https://massbank.eu/MassBank/RecordDisplay.jsp?id=EQ300804&dsn=Eawag

(59)

59

How does PubChemLite perform?

o 111 M => 300 K … how does this influence performance?

MetFragRL, PubChem 2016 MS/MS only (n=473) MetFragRL, PubChem 2016 MS/MS + Metadata (n=1298) MetFragRL, PubChemLite tier0 MS/MS, Ref, Patents, FPSum (n=1298) MetFragRL, PubChemLite tier1 MS/MS, Ref, Patents, FPSum (n=1298)

0 20 40 60 80 100%

(60)

60

NORMAN Suspect List Exchange: Capturing Knowledge

Schymanski et al. (in prep.) https://www.norman-network.com/nds/SLE/

o https://www.norman-network.com/nds/SLE/

o https://zenodo.org/communities/norman-sle

o https://pubchem.ncbi.nlm.nih.gov/classification/#hid=101

Nu mb e r o f e n trie s

PFAS

Pesticides

Pharma

Priority Lists

CCS

Smoke

Dust

Water

Plastics

Surfactants

REACH

Toxins

NPS

>73 lists!

(61)

61

NORMAN-SLE on Zenodo

https://zenodo.org/communities/norman-sle/

(62)

62

NORMAN-SLE on PubChem

Huge thanks to Evan Bolton, Jian Zhang (Jeff), Paul Thiessen, Ben Shoemaker and the PubChem team for this!

(63)

63

NORMAN-SLE on PubChem

o https://pubchem.ncbi.nlm.nih.gov/classification/#hid=101

(64)

64

Missing Entries in PubChemLite

o Assessing the missing entries …

PubChemLite tier0 18 Nov 2019 MS/MS, Ref, Patents, FPSum (n=977) PubChemLite tier0 14 Jan 2020 MS/MS, Ref, Patents, Anno (n=977)

0% 20 40 60 80 100

(65)

65

Transformation Products: Filling the Data Gaps!

(66)

66

NORMAN-SLE on PubChem

(67)

67

Missing Entries in PubChemLite

o Assessing the missing entries …

PubChemLite tier0 18 Nov 2019 MS/MS, Ref, Patents, FPSum (n=977) PubChemLite tier0 14 Jan 2020 MS/MS, Ref, Patents, Anno (n=977) PubChemLite tier0 22 May 2020 MS/MS, Ref, Patents, Anno (n=977) PubChemLite tier0 12 Jun 2020 MS/MS, Ref, Patents, Anno (n=977)

0% 20 40 60 80 100

(68)

68

“Real Life” Application

Jessy Krier (2020) Master’s Thesis, defended July 2020. Figure 34.

(69)

69

Suspect Screening for Lux-Relevant Pesticides

Jessy Krier (2020) Master’s Thesis, defended July 2020. Figure 8

Jessy Krier (2020) S69 | LUXPEST, Zenodo. http://doi.org/10.5281/zenodo.3862689

36 pesticides at Level 2a (MoNA>0.9)

(70)

70

Data-Driven TP/Metabolite Search

Jessy Krier (2020) Master’s Thesis, defended July 2020. Figure 8.

36 pesticides at Level 2a (MoNA>0.9)

(71)

71

Literature Mining for Metabolites / TPs – and Curation

https://www.simolecule.com/cdkdepict/depict.html

(72)

72

Living Data Connections

https://git-r3lab.uni.lu/eci/pubchem/

LCSB-ECI & PubChem Team. DOI 10.5281/zenodo.3890392

(73)

73

Pesticides & TPs over time (tent. IDs)

Jessy Krier (2020) Master’s Thesis, defended July 2020. Modified from Figure 33.

(74)

74

More Examples? Thirdhand Smoke (THS) in Dust

Schymanski, EL, Torres, S, & Ramirez, N. (2020). Thirdhand Smoke in House Dust. Zenodo. http://doi.org/10.5281/zenodo.3613472

+

(75)

75

New MetaData: Disease-Specific Reference Counts

Schymanski et al. DOI:10.1039/C9EM00068B(Perspective) Environ. Sci.: Processes Impacts, 2020.

S37 LitMinedNeuro: DOI: 10.5281/zenodo.3242298

(76)

76

Future Perspectives

ENTACT: Elin Ulrich et al. 2018,

https://www.epa.gov/sites/production/files/2018- 06/documents/comptox_cop_6-28-18.pdfand Ulrich, EM, et al. (2019) Anal Bioanal Chem.

DOI: 10.1007/s00216-018-1435-6 Classification Figures:

Anjana Elapavalore, ECI (Master’s thesis)

o Rapid classification / interpretation of large collections

(77)

77

Next Big Challenge: Connection to Effects!

Modified from Escher, Stapleton, Schymanski (2020). Science. DOI: 10.1126/science.aay6636

(78)

78

Take Home Messages

o Challenge:

Increase % identified

Improve interpretation

More? Check out our presentations at: https://zenodo.org/communities/lcsb-eci/

(79)

79

Take Home Messages

o Challenge:

Increase % identified

Improve interpretation

o “Environmental Cheminformatics” is …

o Capturing expert knowledge

Machine and human readable

o Connecting this to environmental observations

o Identifying and closing knowledge gaps

o Supporting interpretation of complex data

(80)

80

Take Home Messages

o Challenge:

Increase % identified

Improve interpretation

o “Environmental Cheminformatics” is …

o Capturing expert knowledge

Machine and human readable

o Connecting this to environmental observations

o Identifying and closing knowledge gaps

o Supporting interpretation of complex data

o Finally … information in the public domain helps everybody

o You never know when it will help you 

(81)

81

Acknowledgements

Slides available under DOI: 10.5281/zenodo.4075013

emma.schymanski@uni.lu and @ESchymanski

Further Information:

https://pubchem.ncbi.nlm.nih.gov/

https://git-r3lab.uni.lu/eci/shinyscreen

https://git-r3lab.uni.lu/eci/pubchem/

https://ipb-halle.github.io/MetFrag/

https://www.norman-network.com/nds/SLE/

Simone Witzmann, Vero Briche, Rick Helmus, Jessy Krier Todor Kondic, Emma Schymanski, Anjana Elapavalore, Hiba Mohammed Taha; Mira Narayanan, Corey Griffith, Lorenzo Favilli, German Preciat, Adelene Lai, Randolph Singh Evan Bolton,

Paul Thiessen, Jeff (Jian) Zhang,

PubChem Team

Christoph Ruttkies, Steffen Neumann,

MetFrag Team

Philippe Diderich (eau.etat.lu)

(82)

82