LengthLogD: a molecular-size-aware ensemble framework for peptide lipophilicity prediction via multi-scale feature integration

Peptide lipophilicity, commonly measured as logD, is an important physicochemical property associated with membrane permeability, solubility, and bioavailability. Accurate logD prediction remains challenging for structurally complex peptides because available experimental datasets are often limited in size and contain heterogeneous molecular scaffolds. This study presents LengthLogD, a molecular-size-aware ensemble framework for peptide logD prediction based on stratified modelling and multi-scale molecular descriptors. We evaluated LengthLogD on 228 experimentally measured peptide logD records derived from CycPeptMPDB. Molecular representations were constructed from RDKit descriptors, molecular fingerprints, graph-theoretical descriptors, SMILES-derived size information, and Molecular Operating Environment (MOE) descriptors. Records were assigned to three representation-size strata according to SMILES-derived representation length, and category-specific ensemble models based on Ridge regression, Random Forest, and XGBoost were trained and evaluated. In internal evaluation, the fixed-weight ensemble achieved $$R^2$$ values of 0.846, 0.819, and 0.876 for the small, medium, and large representation-size strata, respectively. Ablation and open-source descriptor analyses were used to assess the contribution and reproducibility of different feature groups, including MOE-derived and graph-based descriptors. LengthLogD suggests that stratified modelling and multi-source molecular descriptors can improve internal logD prediction for peptide datasets with heterogeneous molecular complexity. However, the present study is based on a relatively small, assay-specific peptide dataset and requires further validation using larger independent datasets, open-source descriptor sets, and analogue-aware evaluation protocols before broad application in peptide drug discovery.