CNN-BLSTM based deep learning framework for eukaryotic kinome classification: An explainability based approach.

Area of Science:

• Proteomics
• Bioinformatics
• Computational Biology

Background:

• Protein family classification from sequences is crucial in proteomics.
• Current deep learning models lack reliability due to their black-box nature.
• Explainable AI is needed to enhance trust in biological sequence analysis.

Purpose of the Study:

• To develop and validate a novel explainability pipeline for deep learning models in eukaryotic kinome classification.
• To improve the reliability and trustworthiness of AI-driven biological sequence analysis.
• To identify key features influencing kinase classification decisions.

Main Methods:

• Comparative analysis of state-of-the-art deep learning algorithms.
• Selection and application of a CNN-BLSTM model for classifying eight eukaryotic kinase families.
• Integration of GRAD CAM and Integrated Gradient (IG) for model interpretation.
• Experimental validation by masking identified kinase domain traces.

Main Results:

• The CNN-BLSTM model achieved high accuracy in classifying kinase sequences.
• The explainability pipeline successfully identified pivotal kinase domain traces.
• Masking these identified domains caused a significant drop in F1-score (0.96 to 0.76).
• Results align with Explainable AI principles, demonstrating model trustworthiness.

Conclusions:

• The proposed explainability pipeline enhances the trustworthiness of deep learning models for biological sequence classification.
• This approach provides interpretable insights into kinase classification, moving beyond black-box predictions.
• The findings support the broader application of Explainable AI in proteomics and related fields.