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

Sammanfattning

Cyclic peptides are an important class of therapeutic molecules, but their development is oftenlimited by poor membrane permeability and oral bioavailability. Despite recent progress, existingprediction models primarily rely on whole-molecule or graph-based representations and donot explicitly model peptide sequences at the monomer level. Because peptide sequence and localmonomer context determine higher-order structure and physicochemical behavior, monomerlevelsequence modeling provides a natural framework for learning permeability-relevant interactionsin cyclic peptides. Here, we introduce MonoSeqCP, which, to our knowledge, is the firstmodel to predict cyclic peptide membrane permeability using a fully monomer-level, sequencebasedrepresentation. MonoSeqCP is a multimodal transformer that integrates monomer-levelphysicochemical descriptors, extended connectivity fingerprints (ECFP), and explicit connectivityinformation derived from HELM representations. Because non-lariat cyclic peptides do nothave a unique start position, the model explicitly enforces rotational invariance during both trainingand inference to ensure predictions are independent of arbitrary sequence linearization. Themodel was trained and evaluated using the benchmark samples and train–test split reported byLiu et al. Under this protocol, MonoSeqCP showed competitive performance relative to the bestbenchmarking models, achieving an R2 of 0.59 on the held-out test set. When trained on the fullcurated dataset using a random stratified split, the model achieved an R2 of 0.61, demonstratingstrong performance under near–in-distribution conditions. Additional out-of-distribution analysesrevealed substantial performance degradation for peptides containing monomers not observedduring training, particularly when combined with broader distributional shifts, indicatingthat generalization is primarily limited by monomer coverage rather than model architecture.These results establish sequence-level modeling as a powerful framework for cyclic peptide permeabilityprediction, while highlighting the need for broader monomer and chemical coverageto enable reliable extrapolation.