Do deep learning models make a difference in the identification of antimicrobial peptides?

Area of Science:

• Biochemistry
• Computational Biology
• Machine Learning

Background:

• Antimicrobial peptides (AMPs) are explored as alternatives to antibiotics.
• Machine learning models are developed to predict AMP activity.
• Shallow learning models require molecular descriptors, while deep learning models do not.

Purpose of the Study:

• To analyze methodological flaws and biases in deep learning models for AMP prediction.
• To compare the performance of deep learning and shallow learning models for AMP identification.
• To determine if deep learning offers superior prediction capabilities for AMPs.

Main Methods:

• Analysis of pitfalls in deep model development for AMP prediction.
• Fair comparative studies of deep and shallow models on benchmarking datasets.
• Evaluation of model performance in classifying antimicrobial peptides.

Main Results:

• Deep learning models do not outperform shallow learning models in AMP classification.
• Both deep and shallow models encode similar chemical information.
• Predictions from deep and shallow models are highly similar.

Conclusions:

• Deep learning may not be the most suitable approach for AMP identification currently.
• Shallow models offer comparable or superior performance with greater simplicity.
• Deep learning should be considered only when significant performance gains justify the computational cost.