SPLiNet: Lightweight prediction of protein-RNA interaction sites by integrating a structure-aware protein language model with a dual-branch network.

Protein-RNA interactions are central to post-transcriptional regulation, yet residue-level identification of RNA-binding sites remains challenging. Sequence-based predictors often have limited ability to capture long-range dependencies, whereas structure-based pipelines commonly depend on explicit geometric modeling and additional structural preprocessing. Here, we present SPLiNet, a lightweight framework for residue-level protein-RNA binding-site prediction that combines structure-informed protein language model representations with a compact downstream predictor. In the SaProt setting, structural information is introduced through Foldseek 3Di-based tokenization, whereas the downstream predictor does not use explicit coordinates, surface meshes, or geometric graphs. When no reliable structure source is available, we additionally evaluate a fully sequence-only variant based on ESM2. SPLiNet integrates a 1D branch that combines local residue windows with global BiLSTM context and a 2D branch that compresses attention-derived relational patterns. To address severe class imbalance, we evaluate both fixed-threshold and validation-calibrated operating points using temperature scaling, threshold selection, and model ensembling. Across the evaluated benchmarks, the SaProt-based SPLiNet variant generally achieved higher MCC and AUPR than the sequence-only ESM2 variant and remained competitive with representative published baselines. These results support the value of combining structure-informed protein language model representations with lightweight downstream modeling for residue-level protein-RNA interface prediction.