BERTAVP: an interpretable multi-task learning model for identification and functional prediction of antiviral...

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Antiviral peptides (AVPs) show promise for treating viral diseases due to their functional motifs.
• Current models often lack subclass predictions and interpretable motif analysis.
• Challenges include limited peptide representation and imbalanced datasets.

Purpose of the Study:

• To develop an interpretable deep learning framework for identifying AVPs and predicting their functional activities.
• To address data imbalance and peptide representation issues in AVP research.
• To identify key functional motifs responsible for AVP activity.

Main Methods:

• Introduced BERTAVP, a multi-task deep learning framework.
• Utilized BERT and CNN branches for feature extraction (peptide, amino acid, physicochemical).
• Employed focal loss to handle imbalanced datasets and analyzed learned motifs.

Main Results:

• BERTAVP achieved superior performance in AVP identification and functional prediction.
• Identified ELDKWA and SLWNWF motifs as critical structural and functional modules.
• Discovered WMEWDREI (anti-HIV) and KxHxx (coronavirus-specific) motifs within tryptophan-rich regions.

Conclusions:

• BERTAVP provides an effective, interpretable approach for AVP research.
• Motif analysis reveals specific structural elements driving AVP activity.
• The framework enhances understanding and prediction of AVP efficacy against diverse viral threats.