Note: ENNAACT is a novel tool which employs neural networks for anticancer activity classification for therapeutic peptides

Note: ENNAACT is a novel tool which employs neural networks for anticancer activity classification for therapeutic peptides

doi:10.1016/j.biopha.2020.111051

• Using a novel sequence-based deep neural network classifier to predict ACP

• ENNAACT is comparable to best-in-class, CV accuracy ~ 98.3%, Mathews correlation coefficient ~ 0.91, AUC ~ 0.95

• https://research.timmons.eu/ennaact

ACP

• 5-30 AA

• Often drive from host defence peptides (HDP ~ peptide against microbes)

3 mechanisms have been proposed in terms of killing cancer cells

• Cytoplasmic mb disruption via micellization or pore formation

• Induction of apoptosis via disruption of the mitochondrial mb (peptide entering the cells w/o cell mb disruptive?)

Healthy cell vs cancer cells

Healthy cells

• Zwitterionic cell mb

Cancerous cells

• Cell mb ~ net negative charge, this is due to

• phosphatidylserine 

• O-glycosylated mucins

• Sialylated ganglioside

• Heparin sulfate

• High cell surface area

• Increased membrane fluidity

Most previous machine learning method

• SVM

• Random forests

Features used in previous studies;

• Physicochemical features

• AA composition

• Dipeptide composition

• Cho’s  pseudo AA composition

Deep learning approach (part of the ML)

• Using neural network 

ENNAACT

• Using neural networks approach to create a model

• Using primary AA sequence as input to generate model

• https://research.timmons.eu/ennaact

Chou’s 5 steps rule for ML classifier

• Valid dataset

• Effective representation

• ML should be trained on the representation of the dataset

• Predictive performance ~ evaluate through means of CV

• Predictor made available as web-server

NN developed well with large dataset

Source of ACP;

• DBAASP

• dbAMP

• CancerPPD

• BioPepDB

Peptide used as negative (bc no experimental validated for non-ACP activity)

• Random non-secretory sequences from universal protein resource

• ACP ~ usually secretory in nature (based on antimicrobial activity paper)

After cleaning (remove redundancy and gather all information)

• ACP ~ 659 experimentally validate

• non-ACP ~ 5298 peptides

• 7-40 AAs

Model validation

• Dataset is split to 12

• 10 used for 10-fold cross validation

• Model from cross validation -- > ensemble and evaluate through the external dataset (2 remaining subsets)

Feature extraction

• Physicochemical descriptors

• Composition descriptors

ML

• PCA

• t-SNE

• SVM

• RF

• Dense fully connected neural networks

Neural network

• Using keras and tensorflow to build up

Performance evaluation

• Metrics evaluate imbalance dataset

• Acc/Sn/Sp/MCC

• ROC

They manually check the ACP is the real ACP from the original literature.