Learning meaningful representations of protein sequences.

Area of Science:

• Computational Biology
• Machine Learning
• Bioinformatics

Background:

• Data representation significantly impacts information extraction from biological datasets.
• Machine learning models can automatically learn data representations, but sensitivity to model choices affects biological interpretations.
• Defining meaningful representations for biological data, particularly protein sequences, remains a challenge.

Purpose of the Study:

• To investigate optimal data representation strategies for protein sequences in machine learning.
• To evaluate contemporary practices in transfer learning and interpretable learning for biological data.
• To enhance the interpretability of machine learning models for biological sequence analysis.

Main Methods:

• Explored representation learning in the contexts of transfer learning and interpretable learning.
• Assessed the impact of different machine learning model configurations on data representations.
• Incorporated representation geometry into interpretable learning models for protein sequences.

Main Results:

• Identified suboptimal performance in several current transfer learning practices for biological data.
• Demonstrated that incorporating representation geometry significantly improves model interpretability.
• Revealed that geometric considerations enable machine learning models to uncover obscured biological information in protein sequences.

Conclusions:

• The choice of data representation is critical for extracting meaningful biological insights using machine learning.
• Considering representation geometry is key to improving interpretability and uncovering hidden biological information in protein sequence analysis.
• This work offers a pathway to more robust and interpretable machine learning applications in biology.