Increasing resistance toward the conventional antibiotics has become a global concern. Antimicrobial peptides (AMPs) are potential alternatives for conventional antibiotics. Due to cost related reasons in designing and synthesis of AMPs. Machine learning based prediction tools are indispensable.
PREDICTION OF ANTIMICROBIAL PEPTIDES USING MACHINE LEARNING METHODS
1. PREDICTION OF ANTIMICROBIAL
PEPTIDES USING MACHINE LEARNING
METHODS
BILAL NIZAMI
M.Tech (Bioinformatics)
Under the guidance of
Dr. SUSAN THOMAS
Biomedical Informatics Center (BIC)
NIRRH, Mumbai
2. We will be discussing..
• The problem
• The solution
• Objectives
• Literature reviews
• Machine learning in biological problems
• Antimicrobial activity prediction
• Technical background
• Methodology
• Results
• Conclusions
• Future perspective
• Availability and publications
3. The Problem
• Increasing resistance toward the conventional antibiotics has become a
global concern.
Source:-CenterforGlobalDevelopment(CGD)
4. The solution
• Novel antibacterial agents
• Antimicrobial peptides (AMPs) are
potential alternatives for conventional
antibiotics because of-
1.ability to kill target cells rapidly.
2.broad spectrum of activity.
3. and modularity.
5. Yet another obstacle
• Exact MOA and SAR of AMPs is not known completely *.
• Many reasons can be given for the same :-
1. Diversity in AMPs sequence
2. Varied structures
3. Unorganized structure in solution
4. Unknown structure of numerous AMPs.
• Above and beyond high throughput screening, methods for large scale
synthesis and automated assay techniques, two other important pre
requisites are
a) open source in silico libraries of AMPs
b) efficient computational methods.
• A computational method includes prediction tools for antimicrobial
activity.
* Mohammad Rahnamaeian: Antimicrobial peptides Modes of mechanism, modulation of defense
responses, Plant Signaling &Behavior 6:9, 1325-1332; 2011Landes Bioscience.
6. Objectives
• Machine learning based prediction tools for
antimicrobial activity.
• Comparison of SVM, RF and ANN based prediction
models
• Relative importance of various peptide descriptors in
prediction ability of models.
7. Literature Reviews
• AMPs are Abundant and diverse group of biomolecules.
• Selectively lethal against microbes.
• Found every where e.g. Monera(Eubacteria), Protista
(protozoans and algae), Fungi (yeasts), Plantae (plants) and
Animalia (insects, fish, amphibians, reptiles, birds and
mammals). (Sang Y et al. 2008 )
• Exist as α-helical peptides, and β-sheet peptides.
• Difference between cell membrane’s composition,
polarization, and structure of eukaryotes and prokaryotes is
responsible for selective action. Brogden KA (2005)
• Attraction, attachment and pore formation are seen during
the action of AMPs (Roland ,2009)
8. Literature Reviews…
• Two significant properties which are considered for de-novo
design of AMPs (Richard W. 2008 and Prenner 2005)
1. Net positive charge to interact with negatively charged
bacterial membrane.
2. Amphipathic structure to facilitate its integration into the
bacterial membrane. (Sarika P 2011)
Red, basic (positively charged) amino
acids
Green, hydrophobic amino acids
Michael Zasloff (2002)
9. Machine learning in Biological problems
• 1958 - First attempt to model neuronal architecture of the brain.
• 1982 - Stormo et al. proposed ‘Perceptron’ algorithm to distinguish E. coli
translational initiation sequences from other sites.
• machine learning is employed for :-
1. Prediction models
2. Automatic annotation
3. Protein structure and function prediction
4. Active sites determination in proteins
5. Evolutionary analysis
6. Determination of binding sites on protein target
7. Biological network analysis
8. Patterns discovery in biochemical pathways
9. Phylogenetic tree analysis
10. Identifying genetic markers of disease.
10. Antimicrobial activity prediction
• several machine learning based prediction methods have been developed
* support vector machines (SVM), discriminant analysis (DA), Sliding window (SW), artificial neural network
(ANN), quantitative matrix (QM), Hidden markov model (HMM), sequence alignment (SA), Weighted finite-
state transducers (WFST)
• Still a huge gap exists between what need to be achieved and what has
been achieved.
Algorithm / method * Reference Associated database
SVM Lata et al. AntiBP
ANN Lata et al. AntiBP
SW Torrent et al. --
DA Thomas et al. CAMP
QM Lata et al. AntiBP
WFST Whelan et al. --
HMM Hammami et al. PhytAMP
Hammami et al. BACTIBASE
SA Wang et al. APD2
11. Antimicrobial activity prediction
• This is a challenging task, due to
• Low sequence similarity among diverse AMPs (Hancock RE
1999)
• Unorganised conformation
• Moreover costly experimental methods
• So we need good prediction models
• Physicochemical properties like Charge, size, amphipathicity,
amino acid composition, structural conformation,
hydrophobicity and polar angle are responsible for
antimicrobial activity.
• Total of 257 peptide descriptors - which includes dipeptide
and tripeptide composition, composition based on reduced
alphabets, amino acid indices, charge, and hydrophobicity
indices.
12. Technical background
SVMs
• Supervised learning model.
• Originally it was for linearly separable case.
• In 1995 it was extended to the linearly non separable cases
also.
15. Non linear SVM
• Kernel trick.
• Data points are nonlinearly mapped to a feature space of high
dimensions.
• The transformation used is f([x y]) = [x y (x^2+y^2)].
16. Random Forest
• Ensemble learning framework.
• It raises multiple classification trees.
• Decision tree is a common flow chart like schema to represent
classification problems.
17. Random forest..
• Each decision tree in RF is grown as follows :-
• Sample N cases (1/3 of original dataset)with replacement from the original data.
• Select randomly m predictor out of the M predictors (m<<M) and variable that
provides the best split is used to split the node.
• Each tree is grown to its largest possible extent & each tree votes for ‘class labels’.
• The classification winning most votes are chosen.
18. Advantages of RF
• High prediction accuracy.
• Hold perfectly good for large scale dataset with large number
of variables.
• Integral variable selection based on importance and variable
interaction.
• Deals efficiently with data having missing values.
• Ability to reuse forest for future estimation.
• Computation of relation between variables and classification.
• Proximity calculation between cases.
• Can be used for unsupervised learning and outlier detection.
• Internal unbiased estimate of the generalization error
21. Perceptron learning rule
• It involves learning to fix the weight vector so that it is able to predict
correct ±1 output.
• It is a method to alter and re-adjust the weights.
Perceptron rule
• Assign initial weights randomly.
• Then iteratively apply the perceptron.
• If perceptron mis calculate the output, readjust the weights. Repeat this.
Delta rule
• Perceptron rule fails to converge in nonlinearly separable case.
• Based on gradient descent search algorithm.
• Searches the suitable weight from a hypothesis space of weights.
23. Methodology …
• CAMP currently contains 4020
AMPs
• Sequences having X was removed.
• redundant sequence - Cd hit (cut-
off of 0.9)
• Final negative dataset - 4011
sequences.
• Perl script to calculate 257 peptide
features.
• train and test data -70:30.
• Best 64 features - RF Gini score
• Package randomForest in R for RF.
• 1000 tree and default mtry.
• Kernlab package for SVM,
Polynomial kernel.
• nnet package for ANN. Log liner
model with 65 weights.
• Package “ROCR” for evaluation.
24. Results
• 1470 AMP and 532 NAMP in test dataset.
• RF shows the best prediction accuracy
Algorithm MCC against test
dataset
Prediction Accuracy
(in %)
AUC of ROC
curve
RF 0.87 94.2 0.98
SVM 0.82 92.3 0.97
ANN 0.74 87.9 0.94
25. Comparison with other prediction
tools
Server / tools Prediction accuracy (%)
RF SVM ANN SW QM
Our method 94.2 92.3 87.9 -- --
AntiBP -- 92.1 88.17 -- 90.37
AMPA -- -- -- 85 --
Random Forest (RF), Support vector machines (SVM),
Artificial neural network (ANN), sliding window (SW),
quantitative matrix (QM)
26. Fig 1 Fig 2
Fig 3
Figure 1 - Plot of cumulative error
rates in RF - black (overall), red - class
0 (AMP), green - class 1 (NAMP)
Figure 2 - A variable importance plot.
Variable importance is determined by
Mean decrease in Gini score.
Figure 3 - Scatter plot of RF model
(red triangle - AMP and black circle –
NAMP).
27. Conclusions
General conclusion
• Prediction tools are very crucial for designing and synthesis of novel AMPs.
• Sequence of an AMP plays an important role in antimicrobial activity.
• It is necessary to understand the role of peptide feature in antimicrobial
activity.
• Prediction accuracy relies on the relevant information contained within
the descriptors.
Specific conclusions
• RF has higher prediction performance. Ensemble technique seems to be
the reason behind this.
• Best 64 peptide features is identified.
• The prediction tools developed during this study will certainly help in
identifying the new potential AMPs.
28. Future prospective
• Better prediction methods - by incorporating diverse peptide
features & more stringent noise removal strategy.
• Antimicrobial region prediction in a peptide would be very
useful.
• Developing a benchmark dataset would be a great milestone.
• Position specific scoring matrix (PSSM) based prediction.
• Classifying a predicted AMP into further sub families based on
functions. Although this work has been done, it still leaves the
room for improvement in accuracy and methodology.
29. Availability & Publication
• Version 2 of CAMP
http://www.bicnirrh.res.in/antimicrobial/
• Publication of CAMP version 2 is in
communication with Nucleic Acid research
(NAR) http://nar.oxfordjournals.org/.