Research map / Label-efficient and robust learning / Recognition and long-tail

CoPRA: Bridging Cross-domain Pretrained Sequence Models with Complex Structures for Protein-RNA Binding Affinity Prediction

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Multilingual language representation pretraining · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CoPRA, which fuses frozen ESM-2 (protein) and RiNALMo (RNA) interface embeddings with complex structure via a Co-Former pretrained on PRI30k with contrastive and distance-modeling objectives, reaching state-of-the-art protein-RNA binding affinity prediction on the new PRA310/PRA201 datasets and the mCSM mutation set.

Problem

Predicting protein-RNA binding affinity (and mutation-induced affinity changes) from complex structure, where prior methods use only sequence or only structure and labeled complex data is scarce and inconsistent.

Approach

Bridge a protein language model and an RNA language model through the binding-interface complex structure using a dual-path Co-Former with structure-guided attention and outer-product updates, pretrained with bi-scope objectives (cross-modal contrastive interaction CPRI and masked interface distance modeling MIDM).

Key results

• On PRA310 5-fold CV, CoPRA reaches RMSE 1.391, MAE 1.129, PCC 0.580, SCC 0.589, versus 1.446 / 0.522 PCC for the no-pretraining variant and 1.481 / 0.489 PCC for LM+Transformer.
• On the mCSM blind test set after fine-tuning, CoPRA gets RMSE 0.957, MAE 0.833, PCC 0.550, SCC 0.570, outperforming mCSM (PCC 0.528) and DeepNAP.
• Zero-shot CoPRA on mCSM achieves RMSE 0.994 and MAE 0.737, lower than FoldX zero-shot on those error metrics.

Measured effects reported in the paper

Surprising findings

• Pre-trained protein and RNA LMs are kept frozen, and fixed-LM training outperforms LoRA fine-tuning of the LMs in PCC and SCC.
• Prior DeepNAP and PredPRBA perform much worse on PRA201 than originally reported, suggesting limited generalization, possibly because samples overlap their training sets.

Limitations

• Small labeled affinity data (310 complexes) limits generalization.
• Only single-point protein mutations are evaluated; multi-point and RNA mutations are not covered.
• Some baselines (DeepNAP, PredPRBA, PRdeltaGPred, PNAB) were only accessible via web servers, so comparisons are partial.

Experimental setup

Datasets: PRI30k PRA310 PRA201 PDBbind PRBABv2 ProNAB mCSM-NA test set BioLiP2

Benchmarks: PRA310 PRA201 mCSM blind test set

Metrics: RMSE MAE PCC SCC

Baselines: LM+LR LM+RF LM+MLP LM+SVR LM+Transformer DeepNAP PredPRBA FoldX GCN GAT EGNN GVP

Builds on: ESM-2 RiNALMo Invariant Point Attention CLIP AlphaFold2

Closest papers by meaning

• PriFold: Biological Priors Improve RNA Secondary Structure Predictions (AAAI 2025 · similarity 0.86)
• BAnG: Bidirectional Anchored Generation for Conditional RNA Design (ICML 2025 · similarity 0.85)
• Deciphering RNA Secondary Structure Prediction: A Probabilistic K-Rook Matching Perspective (ICML 2024 · similarity 0.85)
• BEACON: Benchmark for Comprehensive RNA Tasks and Language Models (NeurIPS 2024 · similarity 0.85)
• LEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling (NeurIPS 2026 · similarity 0.84)
• Accurate Nucleic Acid-Binding Residue Identification Based Domain-Adaptive Protein Language Model and Explainable Geometric Deep Learning (AAAI 2025 · similarity 0.84)