A to-do list for realizing the sequence-to-function paradigm of proteins

Area of Science:

• Structural biology
• Molecular biophysics
• Computational biology

Background:

• Determining protein function directly from amino acid sequences is a long-standing goal.
• Current homology- or library-based methods struggle with divergent functions from similar sequences.
• The sequence-to-function relationship is linked to protein dynamics, but molecular dynamics data is scarce.

Purpose of the Study:

• To address the underpopulation of molecular dynamics data for machine learning.
• To explore methods for robustly associating protein sequence with function.
• To propose an alternative to computationally intensive, long-term protein dynamics simulations.

Main Methods:

• Literature surveys to identify gaps in machine learning for protein function prediction.
• Developing methods to learn biophysical representations (signatures) from protein dynamics.
• Integrating learned biophysical signatures with existing models.

Main Results:

• Identified a significant gap in molecular dynamics data for machine learning applications.
• Demonstrated the potential of learning biophysical signatures.
• Showcased integrative models combining signatures for improved sequence-function association.

Conclusions:

• Learning biophysical representations offers a viable path to predict protein function from sequence.
• Integrative models enhance the robustness of sequence-function predictions.
• This approach circumvents the need for exhaustive, long-term molecular dynamics computations.