CircRNA binding-site prediction stops reading only the RNA: CPRISP encodes the protein too, reaching mean AUC 0.9874 across 37 benchmarks

When predicting where a circRNA binds an RNA-binding protein, the model now reads multi-level circRNA descriptors and protein-side sequence information together instead of relying on transcript features alone. Earlier deep models emphasized circRNA-derived features, often overlooking nucleotide-level physicochemical and structural information, and frequently underrepresented RBP sequence characteristics that determine binding specificity. Across 37 RBP benchmark datasets, CPRISP reached a mean AUC of 0.9874, against 0.9230 for CircSSNN, 0.8869 for iCircRBP-DHN, 0.8868 for HCRNet, 0.8838 for CRBPDL, 0.8388 for CRIP and 0.8348 for CSCRSites; CPRISP had the highest mean AUC on 32 of the 37.

The circRNA side is split into complementary descriptor channels, the protein side supplies both a global semantic embedding and locally consistent subsequence representations, and the concatenated features go to a convolutional classifier that outputs a binding probability. CircRNA features include k-mer nucleotide frequencies, physicochemical properties such as electron-ion interaction pseudopotential values, six-class secondary-structure context predicted by RNAfold, and Circ2Vec embeddings; each channel is aligned by a convolutional encoder and fused by a multi-scale residual network, while protein sequences use ProteinBERT global features plus random subsequence sampling with contrastive consistency constraints. The CRSP variant without RBP sequence information falls below 0.9 AUC on datasets such as AGO2, EIF4A3 and PTB, and trails CPRISP by more than 0.1 on IGF2BP1 and IGF2BP3; across eight directed cross-cell-line tasks covering four RBPs, CPRISP reached a mean AUC of 0.9221 and ranked first in every task.

The demonstrated scope is 37 circRNA-RBP benchmark datasets, eight directed cross-cell-line tasks for four RBPs, and pathway and compound ranking analyses for nine neurodegeneration-associated RBPs; the authors state that cross-cell-line conclusions are limited to the evaluated proteins and cell-line pairs, and that the compound analysis does not establish direct binding, target engagement or efficacy.

The authors state that SHAP attribution scores should not be read as direct evidence of molecular mechanism, that secondary structure is represented by only six coarse contexts, and that exploratory 3D structure representations did not improve on the RNAfold configuration under the tested distance-histogram encoding; the extent of cross-cell-line generalizability awaits validation across more RBPs and cellular contexts.