PEARL: novel protein targets recommendation model via heterogeneous graph representation learning.

Peptide drug repurposing depends on identifying likely protein targets for a query peptide across a large candidate space, yet most existing peptide-protein interaction models are optimized for isolated pairwise classification rather than target prioritization. Here, we present PEARL, a recommendation-oriented heterogeneous graph learning framework for peptide target recommendation. PEARL represents peptides and proteins as distinct node types connected by known peptide-protein associations and multi-view peptide similarity, and combines graph-based representation learning with similarity-aware score refinement to rank candidate targets for a query peptide. On the UMPPI benchmark, PEARL achieved mean area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPR), and Matthews correlation coefficient values of 0.868, 0.855, and 0.571 in five-fold cross-validation, outperforming the strongest baseline, HGT-PepPI, by 4.4, 4.8, and 7.0 percentage points, respectively. Under clustered cold-start evaluation, PEARL delivered the best overall AUC across the novel-protein, novel-peptide, and novel-pair settings, indicating robust generalization under increasing sequence dissimilarity. PEARL also remained the top performer on two independent test sets, reaching an AUC of 0.863 on Test167 and 0.871 on Test1440. Additional analyses showed that PEARL can be effectively calibrated and that its learned latent space groups functionally related targets beyond simple sequence identity. In a case study of the unseen therapeutic peptide CIGB-814, PEARL ranked SPSB2 as the top candidate target. Docking and molecular dynamics simulations supported a stable binding mode, and surface plasmon resonance further confirmed direct binding with an apparent KD of 5.538 μM. Together, these results support PEARL as a scalable computational framework for peptide target recommendation and peptide drug repurposing.