This work addresses intrinsically disordered proteins (IDPs), whose functions arise from highly heterogeneous conformational ensembles, whereas most existing function prediction methods rely solely on sequence or a single static structure. To bridge this gap, we introduce DynaReason‑Pro—the first framework that jointly predicts Gene Ontology (GO) terms from conformational ensembles and produces natural‑language functional rationales.
We evaluate DynaReason‑Pro on a curated IDP benchmark derived from the Protein Ensemble Database (PED), employing splits that are disjoint both in sequence and structure clusters to avoid data leakage. Moreover, we conduct variability‑stratified evaluation, revealing the most pronounced performance gains in the high‑conformational‑variability regime.
The core pipeline of DynaReason‑Pro consists of:
- Constructing unified feature representations from experimentally or computationally derived multiple conformations;
- Jointly learning GO prediction and text generation via a graph neural network equipped with attention mechanisms;
- Demonstrating the advantage of multi‑conformation information through ablation studies.
Empirical results show that leveraging the full ensemble improves GO prediction F1 scores by roughly 8% and yields significantly more fluent and accurate functional descriptions.
Review