Bacteriocin detection with distributed biological sequence representation

Md Nafiz Hamid, Iddo Friedberg

Bioinformatics and Computational Biology program

Department of Veterinary Microbiology and Preventive Medicine

1

2

Antibiotic Resistance

“A problem so serious that it threatens the achievements

of modern medicine... A post-antibiotic era, in which common infections and

minor injuries can kill, far from being an apocalyptic fantasy,

is instead a very real possibility for the 21st century.”

--WHO (2014)

​

3

Antibiotic Production history

4

Combating Antibiotic Resistance

• Harvest more natural compounds

• narrow spectrum

5

Bacteriocins

• Genetically encoded antimicrobial peptides
• Narrow killing spectrum - ‘designer drug’
• Truly diverse - more potential

Figure: Phylogenetic breadth of bacteriocin killing.

​

KO, Klebsiella oxytoca; KP, Klebsiella pneumoniae; EB, Enterobacter cloacae; CF, Citrobacter freundii; EC, E. coli; SM, Serratia marcescens; HA, Hafnia alvei; VC, Vibrio cholerae

Image from Riley et.al., Annual Review of Microbiology, 2002

Bacteriocin Gene Clusters

6

Figure: Examples of bacteriocin gene clusters

Image from Fields et al., Biochemical Pharmacology, 2016

Context genes

​

​

​

​

​

​

​

​

​

​

​

​

​

7

A

B

T

C

I

P

R

K

F

E

G

Bacteriocin

Immunity

Transporter

Modifier

Regulator

8

A

B

T

C

I

P

R

K

F

E

G

Bacteriocin

Immunity

Transporter

Modifier

Regulator

9

A

B

T

C

I

P

R

K

F

E

G

NisB

NisC

NisT

NisP

Bacteriocin

Immunity

Transporter

Modifier

Regulator

10

A

B

T

C

I

P

R

K

F

E

G

NisB

NisC

NisT

NisP

Bacteriocin

Immunity

Transporter

Modifier

Regulator

11

A

B

T

C

I

P

R

K

F

E

G

NisB

NisC

NisT

NisP

Bacteriocin

Immunity

Transporter

Modifier

Regulator

12

A

B

T

C

I

P

R

K

F

E

G

NisB

NisC

NisT

NisP

Bacteriocin

Immunity

Transporter

Modifier

Regulator

13

A

B

T

C

I

P

R

K

F

E

G

NisB

NisC

NisT

NisP

NisF

NisF

NisE

​

NisG

​

​

NisI

Bacteriocin

Immunity

Transporter

Modifier

Regulator

14

A

B

T

C

I

P

R

K

F

E

G

NisB

NisC

NisT

NisP

NisF

NisF

NisE

​

NisG

​

​

NisI

NisK

NisR

Bacteriocin

Immunity

Transporter

Modifier

Regulator

15

A

B

T

C

I

P

R

K

F

E

G

NisB

NisC

NisT

NisP

NisF

NisF

NisE

​

NisG

​

​

NisI

NisK

NisR

Bacteriocin

Immunity

Transporter

Modifier

Regulator

16

Computational work on detecting Bacteriocins

• BAGEL¹: Database. Also a bacteriocin finding tool that does genome mining with homology based methods.
• Bactibase²: Database. Also offers limited identification facilities.

1: Heel et. al.(2013) 2: Hammami et.al.(2007), 3: Morton et.al.(2015)

• BOA³: Uses context genes to find potential areas for bacteriocins and applies profile hidden markov models.

Multiple sequence alignment

Hidden markov model

17

How BOA works

Jamie Morton

18

How BOA works

19

?

How BOA works

20

Our work - Motivation

• Identification of novel bacteriocins from only primary protein sequences using machine learning classification.

​

?

21

How to represent the protein sequences?

• Distributed representation for each word (in our case - each 3-gram)

​

​

​

​

​

​

MKK =

-0.1211

-0.0381

-0.0228

-0.0833

.

.

.

.

.

.

-0.0945

​

Distributed representation

Figure: 3-grams

MKKAVIVENKGCATCSIGAACGLFGLWG

22

word2vec in Natural Language Processing(NLP)

“You shall know a word by the company it keeps”

-- J.R. Firth

• Goal: Representing a word with a dense vector through training of a Neural Network. Words with similar ‘context’ have similar types of vectors.

• Example: ‘Sophisticated’ and ‘Complicated’. ‘Android’ and ‘iOS’.

Unsupervised Learning

Specially helps when we have small labeled data

23

Cool word2vec effects

Company-CEO

Image from lecture slides of Socher et.al.

24

Cool word2vec effects

Analogies

Image from lecture slides of Socher et.al.

Applications in Natural Language Processing(NLP)

25

• Sentiment classification:
• Positive example: This movie is subtle yet imposing, short yet not hurried.
• Negative example: This movie is short yet taxing, neither subtle nor imposing.

​

Words: This, movie, subtle, yet, imposing, short, hurried, taxing, neither, nor

​

26

How training happens? - Training samples for Neural Network

Corpus: Bacteriocins are much fun and important.

​

Input

Output

Bacteriocins are

​

Bacteriocins much

are Bacteriocins

are much

are fun

Context window: 2

much Bacteriocins

much are

much fun

much and

27

How training happens? - Neural Network

Vocabulary: Bacteriocins, are, much, fun, and, important 6 words

Length: 6

fun

For training sample: (fun, bacteriocins)

bacteriocins

Hidden layer weight matrix

Output layer weight matrix

Dense word vector

The thing we want

From NLP to biology

28

Asgari et.al.(2015)¹ showed efficiency of word2vec in protein family classification.

Asgari et.al., PLOS One, 2015

Figure: t-sne plot for all 3-grams(words) by their biochemical and biophysical properties.

Our word2vec training - Uniprot Trembl

29

MKLIGPTFTNTST

• Corpus: Uniprot Trembl database - 55,899,422 sequences
• Vocabulary: All 3-grams (20 * 20 * 20 = 8000 words)
• Generating 3 sequences from each sequence in the database

3-gram

1st sequence of words: MKL, IGP, TFT, NTS

​

2nd sequence of words: KLI, GPT, FTN, TST

3rd sequence of words: LIG, PTF, TNT

​

1st sequence start

2nd sequence start

3rd sequence start

Our word2vec training - Uniprot Trembl

30

• At the end of training - you get a size 100 dense vector for each 3-gram.

​

​

​

​

​

​

​

​

​

​

MKL =

-0.1211

-0.0381

-0.0228

-0.0833

.

.

.

.

.

.

-0.0945

-0.0145

​

Machine learning classification - Training

31

• 346 positive bacteriocins from BAGEL dataset.
• Each sequence is represented by a size 100 dense vector which is the summation of all the size 100 dense vector of the overlapping 3-grams.

​

​

MKLIGPTFTNTST

MKL + KLI + LIG + IGP + …………………..+TST

-0.1211

-0.0381

-0.0228

-0.0833

.

.

.

.

.

.

-0.0945

-0.0145

​

0.0784

0.0262

0.0529

-0.1338

.

.

.

.

.

.

-0.0988

-0.0381

​

-0.1222 -0.0069 -0.2823

-0.2173

.

.

.

.

.

.

-0.3728

-0.1597

​

0.3240

-0.0407

-0.3316� 0.1062

.

.

.

.

.

.

-0.0260

-0.0282

​

0.2565

-0.0558

0.0221

0.0971

.

.

.

.

.

.

-0.1413

-0.1344

​

+

+

+

+

1.0211

3.6549

-0.3279

-0.9871

.

.

.

.

.

.

5.7712

-3.7610

​

+

+

.

.

.

.

.

.

​

Training - building a negative set

32

• Use the swissprot db - manually reviewed
• CD-HIT to reduce redundancy.
• Get 346 sequences that have the same length distribution as the 346 positive bacteriocins.
• NOT anti-microbial, NOT antibiotic, NOT in plasmid.

Result on the whole dataset

33

Nested 10-fold cross-validation on 692 samples

​

​

Precision =

TP

TP + FP

Recall =

TP

TP + FN

F1 = 2

Precision. Recall

Precision + Recall

ROC curves - SVM

34

Applying trained model on Genbank protein data

35

• 11 high probability bacteriocins from Lactobacillus species
• Collaborators working on experimental validation

Potential new bacteriocins found!!!

36

Lactobacillus acidophilus La-14

Putative bacteriocin

Regulator

Transporter

Regulator

Potential new bacteriocins found!!!

37

Lactobacillus acidophilus 30SC

Putative bacteriocin

Regulator

Regulator

Transporter

Potential new bacteriocins found!!!

38

Lactobacillus acidophilus NCFM

Putative bacteriocin

Transporter

Immunity

Transporter

Transporter

Coming up

39

• Paper
• Software for the community

Thanks!!

40

Iddo Friedberg

Stefan Freed

Shaun Lee

Follow me

​

@naf1zh

Questions!!!

41

42

Lactobacillus acidophilus La-14

Putative bacteriocin

Regulator

Transporter

Regulator

Lactobacillus acidophilus 30SC

Putative bacteriocin

Regulator

Regulator

Transporter

Lactobacillus acidophilus NCFM

Putative bacteriocin

Transporter

Immunity

Transporter

Transporter

43

Lactobacillus helveticus CNRZ32 (Locus: NC_021744, protein id: YP_008235573.1)

​

Putative bacteriocin

Lactobacillus acidophilus NCFM (Locus: NC_006814, protein id: YP_193080.1)

​

Putative bacteriocin

Lactobacillus acidophilus NCFM (Locus: NC_006814, protein id: YP_193019.1)

Putative bacteriocin

44

Lactobacillus acidophilus 30SC

Putative bacteriocin

Regulator

Regulator

Transporter

45

Lactobacillus acidophilus NCFM

Putative bacteriocin

Transporter

Immunity

Transporter

Transporter

46

Bacteriocin

Transporter

Immunity

Regulator

Modifier

Other

Corpus: MKLIGPTFTNTSTLSNSV

​

Input

Output

MKL IGP

​

MKL TFT

IGP MKL

IGP TFT

IGP NTS

Context window: 2

TFT MKL

TFT IGP

TFT NTS

TFT TLS

Corpus: MKL IGP TFT NTS TLS NSV 6 3-grams

Length: 6

MKL

Example training sample: (MKL, IGP)

IGP

Hidden layer weight matrix

Output layer weight matrix

Dense word vector

Corpus: MKL IGP TFT NTS TLS NSV 6 3-grams

Length: 6

MKL

Example training sample: (MKL, IGP)

IGP

Hidden layer weight matrix

Output layer weight matrix

Dense word vector

50

MKLIGPTFTNTST

3-gram

1st sequence of words: MKL, IGP, TFT, NTS

​

2nd sequence of words: KLI, GPT, FTN, TST

3rd sequence of words: LIG, PTF, TNT

​

1st sequence start

2nd sequence start

3rd sequence start

51

​

​

​

​

​

MKL =

-0.1211

-0.0381

-0.0228

-0.0833

.

.

.

.

.

.

-0.0945

-0.0145

​

52

MKLIGPTFTNTST

MKL + KLI + LIG + IGP + …………………..+TST

-0.1211

-0.0381

-0.0228

-0.0833

.

.

.

.

.

.

-0.0945

-0.0145

​

0.0784

0.0262

0.0529

-0.1338

.

.

.

.

.

.

-0.0988

-0.0381

​

-0.1222 -0.0069 -0.2823

-0.2173

.

.

.

.

.

.

-0.3728

-0.1597

​

0.3240

-0.0407

-0.3316� 0.1062

.

.

.

.

.

.

-0.0260

-0.0282

​

0.2565

-0.0558

0.0221

0.0971

.

.

.

.

.

.

-0.1413

-0.1344

​

1.0211

3.6549

-0.3279

-0.9871

.

.

.

.

.

.

5.7712

-3.7610

​

53

Nested 10-fold cross-validation on 692 samples done 50 times

​

​