KDPRA: A Dual-Molecule Knowledge Distillation model with Cross-Attention Fusion for Protein–RNA Binding Affinity Prediction

Siyuan Shen, Zhijian Huang, Yun Shen, Shangqian Wu, Jie chen, Ziyu Fan, Yahan Li, Lei Deng

Keywords: protein–RNA interactions, protein–RNA binding affinity, Bioinformatics

TL;DR: We distill two teachers into a lightweight model for protein–RNA affinity prediction with strong performance on low-resource benchmarks.

Abstract: Quantifying the binding affinity between proteins and RNAs is critical for understanding the recognition mechanisms underlying protein–RNA interactions. However, current computational methods face two major limitations: (1) the scarcity of training data, as experimentally measured protein–RNA binding affinity datasets are limited and insufficient to support the effective training of complex models; and (2) the lack of efficient cross-modal feature interaction mechanisms, which hampers the accurate modeling of the intricate binding patterns between proteins and RNAs. To tackle these challenges, we propose KDPRA, a protein–RNA binding affinity prediction model based on knowledge distillation and a cross-attention mechanism. To better learn residue-level representations of proteins and RNAs, we independently train teacher models for each modality and employ knowledge distillation to guide the student model to learn effective structural and semantic representations. Furthermore, KDPRA incorporates a bidirectional cross-attention fusion module to capture general patterns of protein–RNA interactions. Experimental results demonstrate that KDPRA outperforms existing methods. Case studies further reveal that KDPRA can effectively predict protein–RNA binding affinities, providing strong biological interpretability and promising application potential.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 10997