AMR and machine-learning Prediction of AMR from metagenomes among - - PowerPoint PPT Presentation

AMR and machine-learning

Prediction of AMR from metagenomes among other things

Finlay Maguire finlaymaguire@gmail.com December 3, 2019

Faculty of Computer Science, Dalhousie University

Table of contents

• 1. Genomic Phenotype Prediction
• 2. Non-Bioinformatics Interlude
• 3. AMRtime

1

Genomic Phenotype Prediction

Antibiotic Susceptibility Testing

Bradley et al. (2015)

2

AAFC Salmonella Data-set

3193 (V2) 3125 (G) 3146 (I) 3 1 4 9 ( J 1 ) 3147 (I) 3 1 8 6 ( L ) 3 2 ( N ) 3191 (M) 3 1 9 7 ( M ) 3 1 4 2 ( H ) 3 3 5 3 ( O ) 3 1 3 3 ( H ) 3 1 3 5 ( H ) 3 1 3 7 ( H ) 3 1 3 8 ( H ) 3139 (H) 3156 (K) 3158 (K) 1760 (P1) 1773 (P1) 1 7 6 2 ( P 1 ) 1 7 7 2 ( P 1 ) 1 7 7 ( P 1 ) 1771 (P1) 1 7 6 6 ( P 1 ) 1775 (P1) 1 7 6 7 ( P 1 ) 1 7 6 8 ( P 1 ) 1769 (P1) 3 1 4 ( Q 1 ) 3168 (Q1) 3332 (R) 3342 (R) 1 7 9 2 ( Z ) 1793 (T) 1 8 3 ( A 2 ) 1890 (A2) 1 8 8 8 ( A 2 ) 1 8 9 1 ( A 2 ) 1811 (A2) 3126 (Q2) 3176 (AD) 3128 (AB) 3171 (V1) 3 3 3 9 ( S 1 ) 3 1 4 3 ( A C ) 1 8 9 2 ( A A ) 1893 (AA) 3333 (S2) 3151 (W) 3162 (U) 3 1 9 8 ( X ) 3 1 6 ( U ) 3 1 6 6 ( U ) 3 1 9 9 ( Y ) 1797 (A1) 2003 (B) 2005 (B) 3167 (C) 3169 (D) 3 1 8 ( D ) 3348 (F) 3352 (O) 3 1 8 1 ( E ) 3 3 3 ( F ) 3179 (J2) 3184 (J2) 3305 (S1) 3306 (S1) 3324 (S1) 3 3 5 1 ( S 1 ) 3 3 2 2 ( S 1 ) 3344 (S1) 3 3 4 1 ( S 1 ) 3 3 2 1 ( S 1 ) 3 3 2 6 ( S 1 ) 3132 (Q2) 3314 (S1) 3134 (Q2) 3144 (Q2) 3145 (Q2) 3302 (S1) 3323 (S1) 3311 (S1) 3319 (S1) 3 3 3 6 ( S 1 ) 3337 (S1) 3 3 1 3 ( S 1 ) 3315 (S1) 3318 (S1) 3338 (S2) 3310 (S1) 3349 (S1) 3317 (S1) 1783 (P2) 1 7 5 8 ( P 2 ) 1778 (P2)

3

Genomic RGI Predictions

4

Linking AMR determinants to Phenotype

McArthur et al. (2013)

5

Logistic Regression

RGI =     

amr1 amr2 ... amrJ genome1

1 ... 1

genome2

1 ... 1

...

... ... ... ...

genomeI

... 1      AST =     

abx1 abx2 ... abxK genome1

S S ... R

genome2

R R ... S

...

... ... ... ...

genomeI

S S ... S      βRGI = AST

6

Set-Covering Machines

Genomes AST Decompose into K-mers Genomic K-mers Set-Covering Machine Boolean K-mer Rules

7

AST Prediction Performance

A B C D

A: RGI, B: RGI-efflux, C: Logistic Regression, D: Set Covering Machines. Major Disagreement is overprediction of resistance, Very Major Disagreement is underprediction

8

Learnt features/weights

A B

9

Extending beyond Salmonella

ARO Predictions (Kara Tsang)

10

Extending beyond Salmonella

Logistic Regression

11

Genomic AST Prediction

• Using direct annotations works very poorly across different
• rganisms and resistance mechanisms.

12

Genomic AST Prediction

• Using direct annotations works very poorly across different
• rganisms and resistance mechanisms.
• Even very simple logistic regression models greatly improve

predictions.

12

Genomic AST Prediction

• Using direct annotations works very poorly across different
• rganisms and resistance mechanisms.
• Even very simple logistic regression models greatly improve

predictions.

• Investigation of learnt weights and features can be very scientifically

informative.

12

Non-Bioinformatics Interlude

• Non-profits have data and lots of contextualising knowledge.

13

• Non-profits have data and lots of contextualising knowledge.
• No time or resources to analyse or use it

13

• Non-profits have data and lots of contextualising knowledge.
• No time or resources to analyse or use it
• Informaticians have the skills and resources but no specific

understanding of the context.

13

• Non-profits have data and lots of contextualising knowledge.
• No time or resources to analyse or use it
• Informaticians have the skills and resources but no specific

understanding of the context.

• Many low-hanging fruit that can make big differences.

13

Refugee Women’s Health Clinic

14

Staff Scheduling

15

Language Development in Autism

Qualitative Social Media Analysis (Tamara Sorenson-Duncan)

16

Alpha Diversity of Posting Activity

17

Beta Diversity of Posting Activity

18

Other on-going Projects

• Halifax Community Learning Network
• Shelter Nova Scotia
• 211 Nova Scotia

19

AMRtime

AMR-metagenomics

Genomes Reads AMR Genes

Sequencing AMR detection 20

Why is this difficult?

AMR genes are rare genomically All (~324M) AMR (~2.1M) 107 108

log(Read Count) AMR Reads in Metagenome (0.643%)

2184 CARD-Prevalence Genomes at 1-10X abundance

21

AMR genes have wildly different abundances

1236 AMR PATRIC genomes

22

AMR genes have highly variable diversity

23

AMR sequence space overlaps

1000 500 500 1000 1000 500 500 1000 Actual Families 1000 500 500 1000 1000 500 500 1000 Affinity Clusters (Adj. Rand=0.30041)

MDS of CARD Proteins BLASTP-%ID

24

Insufficient Signal in 250bp Fragments

NDM Multiple Sequence Alignment

25

Insufficient Signal in 250bp Fragments

NDM Multiple Sequence Alignment

26

Other constraints

• No point doing what we do if people can’t use it.
• Limited hardware requirements (a standard workstation or instance

< 8 − 12Gb, 1 − 8 cores).

• Fast enough (< 12 hours).
• Easy to install/configure.
• Easy to use.
• Easy to update.

27

AMRtime

AMRtime structure

Metagenomic Reads Input files Processes Intermediate files Output files AMR Filtering Filtered reads Sensitive Homology Classification CARD Homology predictions Variant Identification Variant predictions Metamodels Metamodel predictions

28

Read filtering

Homology Filter Approaches

10 20 30 40 50

Elapsed Time (hours)

2 4 6 8

Max Resident Memory (GB)

Tool blastn biobloom groot bwa bowtie2 hmmsearch_nt blastx diamond_blastx paladin blastp diamond_blastp hmmsearch_aa

Relative Computational Demands

29

Precision-Recall of Homology Search

0.0 0.2 0.4 0.6 0.8 1.0

Recall

0.2 0.4 0.6 0.8 1.0

Precision

Paradigm BWT BLAST k-mer HMM

30

Optimising for recall

0.90 0.92 0.94 0.96 0.98 1.00

Recall

0.90 0.92 0.94 0.96 0.98 1.00

Precision

Tool blastx bwa diamond_blastx paladin blastp diamond_blastp

31

Sensitive Homology Classification

Dealing with imbalanced training data

Simulated AMR Reads (.fq) Encoding Encoded Reads Stratified Test-Train (20%) Split Labels (.tsv) Training Data Testing Data SMOTE Resampled Training Data Stratified 5-fold CV Training Data Folds

32

What is balance?

• Different gene lengths within families (coverage vs read number)?
• Different family sizes?
• Different family diversity?
• Using a generator to improve on SMOTE.

33

Initial classifier

Training Data Classifier ARO predictions

34

Initial classifier

Training Data Classifier ARO predictions NB 7-mer Average Precision: 0.63

34

Initial classifier

Training Data Classifier ARO predictions NB 7-mer Average Precision: 0.63 %

34

Revised classifier structure: exploiting the ARO

Training Data AMR Family Classifier AMR Families Family 1 SMOTE Family 1 Data Family 1 Classifier Family ... SMOTE Family ... Data Family ... Classifier Family N SMOTE Family N Data Family N Classifier ARO predictions

35

Sequence similarity encoding

Sequence bitscore matrix =       

gene1 gene2 ... genej−1 genej read1

1256 ... 63

read2

...

...

... ... ... ... ...

readi−1

512 ...

readi

... 785 129        Advantages: read length invariant, low dimensionality, uses filtering data computation

36

Cross-Validation

• Encodings:
• Raw sequence
• Filtering homology search family similarity/dissimilarity
• Manual feature extraction (GC/TNF/compositional)
• One-hot K-mer representation
• K-mer embeddings (DNA2vec/BioVec)
• Classifiers:
• Random Forests
• Naive Bayes
• Logistic Regression
• Neural Networks of varying architecture (Torch)

37

Cross-validation Model

0.0 0.2 0.4 0.6 0.8 1.0

Proportion Family Cross-Validation Performance

Metric

Precision Recall

38

Held-out test results

Precision Recall

Family Test Peformance

0.00 0.25 0.50 0.75 1.00

Proportion Normalised Bitscore Random Forest

39

ARO level classification more variable

25 50 75 100 125 150 175 200 225

Ordered AMR Family Index

0.00 0.25 0.50 0.75 1.00

Proportion Median Precision-Recall Within Families

Precision Recall

40

Family diversity as explanation?

100 200 300

AMR Family Cardinality

0.0 0.2 0.4 0.6 0.8 1.0

Precision

41

Within family label imbalance

0.0 0.2 0.4 0.6 0.8 1.0

ARO Proportion of Family Size

0.0 0.2 0.4 0.6 0.8 1.0

Precision

42

On-going Problems

• Multiset prediction when insufficient signal.
• Systematic benchmarking.
• Full end-to-end comparisons with other approaches (soliciting ideas!)
• rRNA and variant models (not discussed here).
• Integration into CARD platform and IRIDA.

43

Acknowledgements

Acknowledgements

• McMaster University: Kara Tsang, Brian Alcock, and Andrew

McArthur

• Simon Fraser University: Fiona Brinkman
• Dalhousie University: Robert Beiko
• Funding: Genome Canada, NERC Undergraduate Student Research

Award, Donald Hill Family Fellowship

44

References

Bradley, P., Gordon, N. C., Walker, T. M., Dunn, L., Heys, S., Huang, B., Earle, S., Pankhurst, L. J., Anson, L., De Cesare, M., et al. (2015). Rapid antibiotic-resistance predictions from genome sequence data for staphylococcus aureus and mycobacterium tuberculosis. Nature communications, 6:10063. Huang, Y., Gilna, P., and Li, W. (2009). Identification of ribosomal rna genes in metagenomic fragments. Bioinformatics, 25(10):1338–1340. Kopylova, E., No´ e, L., and Touzet, H. (2012). Sortmerna: fast and accurate filtering of ribosomal rnas in metatranscriptomic data. Bioinformatics, 28(24):3211–3217.

45

McArthur, A. G., Waglechner, N., Nizam, F., Yan, A., Azad, M. A., Baylay, A. J., Bhullar, K., Canova, M. J., De Pascale, G., Ejim, L., et al. (2013). The comprehensive antibiotic resistance database. Antimicrobial agents and chemotherapy, 57(7):3348–3357. Schmieder, R., Lim, Y. W., and Edwards, R. (2011). Identification and removal of ribosomal rna sequences from metatranscriptomes. Bioinformatics, 28(3):433–435.

46

Backup

46

Variant Models

Ribosomal Variant Models

Metagenomic reads Ribosomal fragment identification 5S//16S/23S binned reads Taxonomic classification Taxa binnned reads Alignment SNPs

47

Identifying Ribosomal Reads

• MetaRNA (Huang et al., 2009)
• Ribopicker (Schmieder et al., 2011)
• SortmeRNA (Kopylova et al., 2012)
• 77 models
• Reads simulated from the underlying 30 species reference genomes

48

Identifying Ribosomal Reads

49

Identifying Ribosomal Reads

50

Identifying Ribosomal Reads

51

Identifying Taxonomy

52

Some are relatively easy

53

Others are a mess

54

Some are group ambiguous

Probably a Mycobacterium?

55

Others are just a toss-up

56

Ambiguity in classification

57

Meta-models

Meta-models

• Efflux Pump
• Gene Cluster

Predicted components Phylogenetic Placement Phylogenetic groups Sequence feature extraction Sequence features Grouping classifier Multicomponent AMR hits CARDPredicted Phylogenies

58