Interpretable molecular encodings and representations for machine learning tasks.

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Machine learning models are increasingly used in biomedical applications, especially for classifying peptides and proteins.
• Current molecular encoding methods often lack the structure required for optimal machine learning model performance.

Purpose of the Study:

• To introduce a novel, interpretable molecular encoding method called Interpretable Carbon-based Array of Neighborhoods (iCAN).
• To improve the performance and interpretability of machine learning models in peptide and protein classification.

Main Methods:

• The iCAN method captures carbon atom neighborhoods in a counting array, creating structured molecular encodings.
• The method enables comparison of molecular neighborhoods, pattern identification, and visualization of relevance heat maps.
• iCAN was applied to peptide classification and extended to protein classification tasks.

Main Results:

• iCAN outperformed a predecessor encoding method in a peptide classification study.
• When applied to proteins, iCAN surpassed a leading structure-based encoding on 71% of datasets.
• The method demonstrated versatility across various organic molecules, including peptides and proteins.

Conclusions:

• iCAN provides interpretable molecular encodings suitable for diverse machine learning applications in biomedicine.
• This method offers a promising new direction for peptide and protein classification, potentially accelerating drug discovery and disease diagnosis.