Rigidity-Aware Geometric Pretraining for Protein Design and Conformational Ensembles

RigidSSL pretraining enhances protein designability by up to 43% and improves conformational ensemble modeling in G protein-coupled receptors.

• Why it matters: Current protein generative models struggle to jointly learn geometry and design, rely on local atomic representations, and fail to capture dynamic conformational information, limiting their effectiveness.
• What they did: The authors developed RigidSSL, a two-phase geometric pretraining framework using 432K structures and 1.3K molecular dynamics trajectories, with a rigidity-aware flow matching objective to optimize conformational dynamics.
• The result: RigidSSL variants significantly boost design success rates, diversity, and novelty in protein generation, and better model biophysically realistic conformations, advancing protein design and structural modeling.