Data-Driven Models for Predicting Intrinsically Disordered Protein Polymer Physics Directly from Composition or...

Area of Science:

• Biophysics
• Computational Biology
• Protein Science

Background:

• Intrinsically disordered proteins (IDPs) present significant challenges for experimental and computational characterization.
• Current protein folding prediction tools like AlphaFold struggle with the dynamic structural ensembles of IDPs.

Purpose of the Study:

• Develop novel representations for intrinsically disordered proteins (IDPs).
• Integrate these representations with machine learning models to predict IDP properties, specifically the radius of gyration and scaling exponent.

Main Methods:

• Introduced a new physically-motivated feature: the "bag of amino acid interactions" representation.
• This feature explicitly encodes pairwise non-bonded interactions within protein sequences.
• Tested this feature alongside traditional methods on a large dataset of coarse-grained simulated IDP sequences.

Main Results:

• The "bag of amino acid interactions" representation significantly outperformed existing categorical and physically-motivated features.
• The new feature demonstrated robust extrapolation capabilities on unseen data.
• The approach showed promise for predicting the radius of gyration and derived scaling exponent of IDPs.

Conclusions:

• The developed representation offers a promising advancement for the computational study of IDPs.
• This method has the potential to provide deeper physical insights into amino acid interactions and their temperature dependence.
• The approach is adaptable for broader applications within protein science.