MonoSeqCP - A multimodal transformer-based model for cyclic peptide membrane permeability prediction

Abstract

Cyclic peptides are an important class of therapeutic molecules, but their development is often limited by poor membrane permeability and oral bioavailability. Despite recent progress, existing prediction models primarily rely on whole-molecule or graph-based representations and do not explicitly model peptide sequences at the monomer level. Because peptide sequence and local monomer context determine higher-order structure and physicochemical behavior, monomerlevel sequence modeling provides a natural framework for learning permeability-relevant interactions in cyclic peptides. Here, we introduce MonoSeqCP, which, to our knowledge, is the first model to predict cyclic peptide membrane permeability using a fully monomer-level, sequencebased representation. MonoSeqCP is a multimodal transformer that integrates monomer-level physicochemical descriptors, extended connectivity fingerprints (ECFP), and explicit connectivity information derived from HELM representations. Because non-lariat cyclic peptides do not have a unique start position, the model explicitly enforces rotational invariance during both training and inference to ensure predictions are independent of arbitrary sequence linearization. The model was trained and evaluated using the benchmark samples and train–test split reported by Liu et al. Under this protocol, MonoSeqCP showed competitive performance relative to the best benchmarking models, achieving an R2 of 0.59 on the held-out test set. When trained on the full curated dataset using a random stratified split, the model achieved an R2 of 0.61, demonstrating strong performance under near–in-distribution conditions. Additional out-of-distribution analyses revealed substantial performance degradation for peptides containing monomers not observed during training, particularly when combined with broader distributional shifts, indicating that generalization is primarily limited by monomer coverage rather than model architecture. These results establish sequence-level modeling as a powerful framework for cyclic peptide permeability prediction, while highlighting the need for broader monomer and chemical coverage to enable reliable extrapolation.