A hybrid deep learning framework for WT or mutant peptide prediction using p53 mutation data

Area of Science:

• Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• The p53 protein is crucial for maintaining genome integrity and acts as a tumor suppressor.
• Single amino acid variations (SAVs), particularly hotspot mutations, can disrupt p53's DNA-binding and transcriptional functions, promoting oncogenesis.
• Understanding the functional impact of p53 SAVs is vital for elucidating disease mechanisms and cellular processes like apoptosis and DNA repair.

Purpose of the Study:

• To develop a computational framework for identifying deleterious p53 SAVs.
• To analyze the structural, sequential, and biochemical effects of p53 variants.
• To provide an interpretable and efficient tool for prioritizing high-risk p53 mutations.

Main Methods:

• Utilized quantitative biochemical descriptors and sequence data of p53 peptides.
• Employed a deep learning architecture combining Convolutional Neural Networks (CNN) and bidirectional Long Short-Term Memory (Bi-LSTM) with attention mechanisms and ESM-2 embeddings.
• Performed systematic analysis using molecular weight, instability index, hydrophobicity, motif enrichment, and amino acid substitution patterns.

Main Results:

• Achieved high predictive performance with 0.98 accuracy, 0.98 AUROC, and 0.99 precision-recall AUC using stratified 5-fold cross-validation.
• Identified key amino acid residues and biochemical factors influencing p53 variant pathogenicity via SHAP-based interpretation.
• Demonstrated the model's robustness in classifying wild-type versus mutant p53 sequences.

Conclusions:

• The developed model offers a robust and interpretable framework for assessing the functional consequences of p53 variants.
• Provides insights into the biochemical and structural impacts of SAVs on p53.
• Presents a computationally efficient tool for prioritizing p53 variants with high oncogenic risk.