Integrating Multi-View Residue Graph and Protein Language Model for Cell-Penetrating Peptide Prediction via Global-Local Graph Aggregation and Cross-Attentive Fusion.

Abstract

Cell-penetrating peptides (CPPs) are promising carriers for intracellular delivery, but large-scale experimental discovery remains costly and slow. Existing computational predictors often rely on handcrafted sequence descriptors and lack biophysical grounding and interpretability. In this study, we present DeepCPP, a dual-branch frame work that integrates a biophysically informed multi-view residue graph with ESM-2 protein language model embed dings. Specifically, the graph branch encodes local sequence context and residue-wise physicochemical similarity via global-local aggregation with top-k subgraph focusing and gated broadcast, while the ESM branch provides context-aware representations. The two views are aligned and fused by cross-attention with an HSIC-based decor relation term, and a Kolmogorov-Arnold Network (KAN) head enhances nonlinear separability. We also curate a newbenchmark from CPPsite3and adoptcluster-controlled splits to reduce leakage and enable credible generalization. Comprehensive evaluations show that DeepCPP outper forms state-of-the-art CPP and peptide-function prediction methods. Interpretability analyses highlight charge clustering and termini patterns, and reveal residue connectivity consistent with oriented amphipathicity, offering actionable guidance for rational design. Overall, DeepCPP provides an accurate, interpretable, and scalable pre-screening tool for CPP discovery.