End-to-End Prediction of Subcutaneous Peptide Absorption Rates from Amino Acid Sequences Using Physics-Informed Machine Learning

Abstract The subcutaneous (SC) absorption rate constant (ka) dictates critical pharmacokinetic and dosing parameters of injectable peptide therapeutics, yet currently, it cannot be predicted prior to clinical testing, forcing empirical decisions at the discovery and formulation stages. We address this important gap by constructing the first end-to-end machine learning framework that predicts SC ka, both point predictions and prediction intervals, purely from the amino acid sequence and conjugate chemistry. For developing the model, we first curated a comprehensive database of 41 human clinical ka values (0.0077−7.61 h−1), covering unmodified, fatty-acid-acylated, and protein-conjugated peptides. These data were subsequently encoded into a combination of protein language model (ESM-2) embeddings, chemical language model (ChemBERTa) embeddings, and 30 physics-informed molecular descriptors designed to capture the physical variables governing SC transport. Seven machine learning (ML) architectures and ensembles thereof were trained, with the best model (NGBoost) achieving R2 = 0.850 under leave-one-out cross-validation, a markedly high accuracy given the large variance in the clinically reported ka values. Notably, models trained on the physics-informed molecular descriptors or combinations thereof substantially outperformed models trained purely on learned embeddings (ESM-2, ChemBERTa)a direct consequence of the small scale of our training dataset. SHAP interpretability analysis indicates that predictions are consistent with established SC absorption physics, with properties including molecular weight, net charge, hydrophobicity, and acyl chain molecular characteristics dictating model predictions. Leveraging Jackknife+, the developed framework also reports ka prediction uncertainties (prediction intervals). Overall, this framework enables high-throughput sequence-level candidate ranking before in vivo data become available, supporting the rational early-stage development of peptide therapeutics.