Generalizable RNA 3D Structure Prediction with RNA-Adapted Protenix and Agentic Tree Search

Predicting RNA three-dimensional structure remains challenging because of limited experimental data, conformational flexibility, and reduced accuracy on long and topologically complex targets. We adapt Protenix for RNA-focused structure prediction using nucleotide-specific inputs, RNA structural supervision, and multi-conformation training. Our approach has three main components. First, multi-conformation supervision retains up to five experimentally observed reference conformations per RNA sequence and optimizes primarily against the best-matching valid state while retaining an auxiliary reference loss. Second, we augment the curated experimental corpus with quality-filtered synthetic RNA structures generated using a diffusion-inspired structural generation pipeline. Third, we use Agentic Tree Search to select five structures from 15 candidates generated by RNA-adapted Protenix, template-based matching, and Boltz, balancing source-model reliability, conformational diversity, and structural priors. RNA-adapted Protenix is trained with MMseqs2-derived evolutionary information, while benchmark inference is performed without inference-time MSA under the competition environment. The strongest standalone configuration reaches TM-scores of 0.489 on the public benchmark and 0.463 on the hidden private benchmark of the Stanford RNA 3D Folding Challenge. Agentic Tree Search increases performance to 0.684 public and 0.632 private. Synthetic augmentation improves the public score further but does not improve the private score, highlighting the importance of evaluating structural generalization on hidden RNA targets.