Researchers have developed ZeroFold, a machine learning model capable of predicting how strongly proteins bind to RNA molecules using only sequence information — no experimentally determined or computationally predicted molecular structures required. The work, posted as a preprint on arXiv (https://arxiv.org/abs/2603.23583) and not yet peer-reviewed, addresses a long-standing challenge in structural biology with implications for gene regulation research and RNA-targeting drug development.

The central problem the team tackled is RNA's inherent flexibility. Unlike proteins, RNA molecules do not adopt a single stable shape; instead, they exist as shifting ensembles of conformations. Most existing prediction tools commit to one predicted structure, discarding potentially critical binding information in the process. ZeroFold sidesteps this limitation by using what the authors call 'pre-structural embeddings' — internal mathematical representations extracted from a biomolecular AI foundation model called Boltz-2, captured at a processing stage before the model commits to a specific three-dimensional structure. These intermediate representations implicitly capture information about the range of shapes an RNA molecule might adopt.

ZeroFold combines these embeddings for both the protein and RNA components through a cross-modal attention mechanism — a technique that allows the model to weigh relationships between the two molecular partners — to produce a binding affinity prediction directly from their amino acid and nucleotide sequences.

To train and test the model, the researchers assembled PRADB, a curated dataset of 2,621 unique protein-RNA pairs with laboratory-measured binding affinities drawn from four existing databases. On a held-out test set designed to avoid overlap with training data using a 40% sequence identity cutoff, ZeroFold achieved a Spearman correlation of 0.65 — a statistical measure of how well its rankings of binding strength matched experimental rankings, where 1.0 would be perfect agreement. The authors argue this figure approaches the practical ceiling set by noise inherent in experimental measurement methods.

When evaluated under increasingly stringent conditions designed to minimize any advantage from similarity to training data, ZeroFold compared favorably against leading structure-based and sequence-based competitor methods, with its relative advantage growing as sequence similarity to competitors' training sets decreased.

Important limitations apply. This is a preprint and has not undergone peer review. The PRADB dataset, while curated from four databases, contains 2,621 pairs — a relatively modest number for training deep learning models, and the authors acknowledge performance could vary for novel protein-RNA combinations far outside the training distribution. The Spearman correlation of 0.65, while competitive, also means meaningful prediction error remains. The model's performance on RNA types underrepresented in existing affinity databases is also unclear.

The researchers argue their approach opens a route to affinity prediction for protein-RNA pairs for which no structural data exist, a category that represents the majority of biologically and therapeutically relevant interactions.