Predictions of antibiotic resistance genes in the human intestinal microbiota

1

Journée des bioinformaticiens de Jouy-en-Josas February 5, 2014

Etienne Ruppé

2

Not considered as ARDs: gene(s), which, when mutated, cause a decrease in the susceptibility to at least one antibiotic (e.g. mutated DNA gyrases cause fluroquinolone resistance)

Adapted from Levy SB, et al. 2004. Nat Med. 10:S122-129.

Antibiotic efflux pump

Chromosome

Antibiotic- altering enzyme

Antibiotic-degrading enzyme

Antibiotic-unsusceptible

metabolic pathway

Antibiotic

Antibiotic

Antibiotic Mobile genetic element

Antibiotic

Antibiotic Target site protection

Considered ARDs: gene(s) which, when expressed, cause a decrease in the susceptibility to at least one antibiotic (e.g. beta-lactamases hydrolyze beta-lactams)

X Mutation X

Porin X Antibiotic

What do we mean by antibiotic resistance determinants (ARDs)?

3

Any ARDs in the intestinal microbiota?

Only a few hundreds of ARDs among millions of genes? > Issue of the ARD database (ARDB not curated) > Issue of the searching method

Forslund, K. et al.(2013) Genome Res; Ghosh, TS. et al. (2013) PloS One; Hu, Y. et al. (2013) Nat Comm

Colonne1 Forslund, K. et al Ghosh, TS. Et al Hu, Y. et al.

Journal Genome Research Plos One Nature Communications

Published in 2013 2013 2013

N individuals 252 257 162

Origin of individuals American, Danish, Spanish American, Danish,

French, Italian, Japanese,

Spanish, Indian, Chinese

Danish, Spanish, Chinese

ARD reference database ARDB (enriched in-house) ARDB ARDB

Search algorithm Blastn Blastx Blastp

N unique ARDs (>95%) 100 157 156

Beta-lactamases TEM, SHV, AmpC_E. coli,

CCRA, CBLA, CFXA, CEPA

TEM, LEN, SHV, OXY,CTX-

M, CFXA, CBLA, CEPA,

AmpC_E. coli, CMY-2

KPC, ROB, TEM, CTX-M, OXY,

PER, SHV, CARB, PSE, LCR, OXA-

1, OXA, SME, L1, IMP

4

1Sommer, M.O., et al. Science 325, 1128-1131 (2009).

Aminoacid identity (NCBI)

METAGENOME

AEROBIC CULTURE

BOTH

A hunch from functional metagenomics… ARDs from the intestinal microbiota might differ from

those in the databases

Non-cultivable under the radar

Cultivable ARD databases

If one-dimension cannot make it, try 3-D

5

Indeed, proteins with low shared identity can have the same structure

PER-1 and CTX-M-15 do not share more than 21% identity in aminoacid (1-dimension)

1 15710 20 30 40 50 60 70 80 90 100 110 120 130 140(1)

MVKKSLRQFTLMATATVTLLLGSVPLYAQTADVQQKLAELERQSGGRLGVALINTADNSQILYRADERFAMCSTSKVMAAAAVLKKSESEPNLLNQRVEIKKSDLVN--YNPIAEKHVN--GTMSLAELSAAALQYSDNVAMNKLIAHVGGPASVTACTX_M_15_JQ686199 (1)

--MNVIIKAVVTASTLLMVSFSSFETSAQSPLLKEQIESIVIGKKATVGVAVWGPDDLEPLLINPFEKFPMQSVFKLHLAMLVLHQVDQGKLDLNQTVIVNRAKVLQNTWAPIMKAYQGDEFSVPVQQLLQYSVSHSDNVACDLLFELVGGPAALHDPER_1_EF5356 (1)

I L AS L L S AQS L I I A LGVAL D IL EKF M S KL A VL D LNQ V I KA LLN W PI H SM L L AL HSDNVA L VGGPAAL Consensus (1)

But they do share a high homology in 3-dimensions!

CTX-M-15 PER-1 Alignement of CTX-M-15 and PER-1 (TMscore > 0.9)

6

Homology modeling

MNIIDIVAIIPYFITLGAT

Template alignment

Backbone generation

Loop modeling

Side chain modeling

Model optimization

7

8

So, why not using 3D instead of 1D to identify ARDs in the interstinal microbiota?

ISSUES

1. Which 3D scores/cut-offs would be used to identify ARDS among candidates?

2. Online tools (Itasser) do not allow high-throughput: one model per 24h, thousands of candidates

3. 3D modeling requires important computational ressources

SOLUTIONS

1. Use a pairwise comparative process

2. Uses of a software that performs fast and accurate modeling

3. Use local/national clusters

9

ARD Reference sequences

Find sequence homologues: Hmmer,

Blastp, SSearch

Protein catalogue

Filtering candidates: size (75%<X<125% of reference mean size)

ARD Reference templates

Negative reference templates

Homology modeling of candidates

Negative reference

sequences

Candidates vs negative templates

Candidates vs ARD

reference templates

Score analysis

Concept of pairwise comparative modeling

PREDICTION

Positive reference

sequences

Functional annotation with eggNOG v3

Substraction of the COG/NOG of the

positive reference sequences

Candidates

Selection of the ‘leading’ proteins of

the remaining groups

Curation

Negative reference sequences

Selection of the negative references 10

Positive and negative

references

Homology modelling with ARD reference

Homology modelling with ref negative

reference

Structural model candidate vs

reference

Structural model candidate vs negative

reference

Model construction with logistic regression and machine learning1

Universal model construction

11000 bootstraps, 10 fold cross-validations

11

Statistical model: high discrimination between ARDs and negative references

AUC = 0.98 662 ARD references 522 negative references

12

Statistical model 13

14

An example of prediction: class A beta-lactamase

TEM (pdb 1TEM)

Candidate from Akkermansia (36% identity with VEB-1) 86% predicted as class A beta-lactamase

Blaa reference templates Negative reference templates

15

Qnr (PDB 2W7Z)

Candidate from Enterococcus (21% identity with QnrA4) 95% predicted as Qnr

An example of prediction: Qnr

Qnr reference templates Negative reference templates

Prediction score and identity with reference ARD

16

17

Validation with an external dataset

1Forsberg, KJ., et al. Nature, 29; 509 (2014), 2Gibson, MK., et al. ISME J. 1;207 (2015), 3The same ARD families as those considered in the PCM and the conventional method were searched. Candidates outside the size ranges were discarded.

12,904 ORFs

Functional metagenomic study from soils: 4654 inserts

containing ARD1

1380 (99.2%) predicted as

ARDs

Flat search: 1674 hits

1391 candidates for PCM

210 from inserts with no identified

ARDs

73 from inserts with

>1 annotated ARDs

1391 candidates for conventional method

(80% identity)

9 (predicted as ARDs

Resfams2

1345 predicted as ARDs3

18

Worth a second round?

Class A beta-lactamases: 204

predictions added to the reference dataset

New PCM round

1334 candidates, 933 after removing 1st

round candidates

70 predicted as class A beta-lactamases

863 not predicted as class A beta-

lactamases

TEM-1 (green) aligned with a blaa obtained during the 2nd round

19

Influence of the size

0

500

1000

1500

2000

2500

3000

3500

Total number of PCM

Candidates

Predicted as ARDs

Summary of the predictions for December 2014

20

~18,000 PCM performed (36,000 structures) ~2 year full time on a 8-CPU cluster Partners involved: -The local MetaGenoPolis cluster -The Jouy-en-Josas INRA cluster (Migale) -The Toulouse INRA cluster (Genotoul) -The Roscoff INRA cluster -The Genouest cluster

21

Do people cluster according to their resistome? Concept of ‘resistotypes’

Ding & Schloss. Nature. 2014 May 15;509(7500)

1 2 3 4 5 6 Resistotype

22

Resistotypes are linked to gene richness, such as enterotypes are

23

Resistotype 4 is related to a low gene richness

Resistotypes are not associated to age, gender, or body mass index

MetaHIT subjects ARDs

24

RLQ ordination method. Dray, S., et al. Ecology. 2014 Jan;95(1):14-21.

aac2 aac6 ant

blaa

aph

blab1

blab3 blac blad qnr

van

Before ATB Right after ATB At distance from ATB

Preliminary results from Utrecht 25

What is the interplay between the resistome and the effect of antibiotics on

the intestinal microbiota?

Resistotype 2 Resistotype 1

Resistotype 3 Resistotype 4 Resistotype 5

Antibiotic exposure

Altered microbiota

26

Resistotype 6

Some commensal bacteria indeed provide a protection against antibiotics

Stiefel, U. et al. 2014 Aug;58(8):4535-42

27

28

Perspectives

What if ARDs from the dominant microbiota were eventually good to us??? If so, what are the genes/species that exert this altruistic effect?

Data from EvoTAR cohorts shall contain all the answers!

Thank you for your attention!

Resistant commensals can protect from colonization by exogenous pathogens/resistant bacteria

They have barely been isolated in pathogens despite close contact and ATB pressure