Pre-Structural Embeddings for Protein-RNA Binding Affinity Prediction

Many biologically important molecules, including RNA and intrinsically disordered proteins, populate conformational ensembles that cannot be fully represented by a single static structure. We propose that intermediate representations from biomolecular foundation models contain rich pre-structural embeddings that can be transferred to downstream tasks involving conformationally flexible biomolecules. We validate this representation strategy on protein-RNA binding affinity prediction and introduce ZeroFold, a transformer-based model that integrates pre-structural embeddings from Boltz-2 through pair-biased self-attention. To ensure rigorous evaluation, we construct PRADB, a curated benchmark of 2,582 unique protein-RNA pairs with experimentally measured affinities, split at a strict 40\% sequence identity threshold for both modalities. On the held-out test set, ZeroFold achieves a Spearman correlation of 0.64 and outperforms state-of-the-art predictors. Ablations support pre-structural embeddings as an optimal representation transfer strategy for learning the interactions of flexible biomolecules from more rigid structures.