CPPainter: Painting Cell-Penetrating Peptides for Identification and Generation through Reinforcement Learning

Cell-penetrating peptides (CPPs) are promising intracellular delivery vectors, but computational CPP identification, sequence optimization, and candidate generation are often studied as separate tasks. This study aimed to develop an integrated framework for CPP prediction and reward-guided sequence design. We developed CPPainter, a computational framework that couples a protein language model ensemble classifier with a reinforcement-learning-guided generator. CPPainter-Classifier outputs a continuous CPP probability score for CPP identification and sequence-level reward assignment. The CPPainter-Generator selects mutation sites with a trainable policy network, proposes amino acid substitutions using a masked language model, and updates the policy through REINFORCE. CPPainter-Classifier achieved a Matthews correlation coefficient (MCC) of 0.9482 on the internal CPPainter-Bench and scores of 0.9002, 0.9348, and 0.7339 on the external CPPCGM-Bench, SiameseCPP-Bench, and MLCPP2-Bench, respectively, outperforming previously reported methods by 1.2 to 7.0 percentage points. Starting from three seed pools, including known CPPs, bacterial signal peptides, and random sequences, the CPPainter-Generator produced 924 high-confidence, strictly novel CPP candidates. The generated candidates successfully reproduced canonical CPP amino acid patterns and physicochemical profiles, achieving Pearson correlations against the positive reference of 0.97 for molecular weight, 0.86 for the Eisenberg hydrophobic moment, and 0.80 for GRAVY hydropathy. CPPainter provides a unified computational framework for CPP identification, sequence optimization, and candidate generation. The generated sequences should be regarded as computationally prioritized CPP candidates and require further experimental validation.