Synthetic sequence alignments as programmable probes of learned conformational landscapes in deep learning protein...

Area of Science:

• Structural biology
• Computational biology
• Deep learning in protein science

Background:

• Protein conformational flexibility is crucial for function, but predicting alternative states is challenging.
• Deep learning models (AlphaFold2, AlphaFold3, RoseTTAFold2) predict static protein structures but their understanding of conformational landscapes is unclear.

Purpose of the Study:

• To develop a method for interrogating the internal logic of deep learning protein structure prediction systems.
• To systematically bias prediction models toward specific conformational states using designed inputs.
• To understand how these models weigh different sources of information (sequence vs. alignment).

Main Methods:

• Introduction of synthetic multiple sequence alignments (MSAs) designed via inverse folding.
• Systematic biasing of AlphaFold2, AlphaFold3, and RoseTTAFold2 using synthetic MSAs.
• Adversarial experiments pairing query sequences with competing MSAs.
• Initialization of predictions from molecular dynamics trajectories.
• Use of hybrid alignments combining synthetic and natural MSA segments.

Main Results:

• Synthetic MSAs successfully bias prediction models toward distinct conformational states, including those not accessible via natural sequences.
• Adversarial experiments reveal sequence-dependent responses and expose how alignment-derived and sequence-derived signals are weighted.
• Predictions show a bias towards compact, training-distribution-favored conformations.
• Hybrid alignments allow targeted steering toward specific conformational states.

Conclusions:

• Synthetic MSAs provide a generalizable framework for dissecting conformational landscapes encoded by deep learning predictors.
• This approach enhances understanding of model behavior and enables access to biologically relevant hidden protein states.