Monoclonal Antibody Developability from Early Assay Panels: Machine Learning for Formulation and Pharmacokinetic Risk

Abstract

Formulation and pharmacokinetic liabilities remain major bottlenecks in monoclonal antibody development. Here, we present the agnostic context-embedding transformer (ACeT), an interpretable machine-learning framework that integrates heterogeneous early assay panels into endpoint-specific developability models. Using published antibody datasets, ACeT predicted high-concentration viscosity from four dilute-solution assays with held-out R{superscript 2} {approx} 0.75 and root-mean-square error {approx} 4.8 cP across a 5-45 cP range. From four clearance-related in vitro assays, it predicted mouse intravenous exposure with held-out R{superscript 2} = 0.80 and normalized root-mean-square error = 0.15. On a public 152-antibody panel, ACeT predicted hydrophobic interaction chromatography retention time, an orthogonal stickiness/hydrophobicity readout, with out-of-fold Pearson r{superscript 2} {approx} 0.83 and outperformed a published quantitative structure-property relationship baseline. In an exploratory retrospective analysis using five early developability assays, ACeT classified clinical outcomes (Approved vs Terminated) with balanced accuracy of [~]0.78 on a held-out internal set of 23 clinical IgG1 antibodies with outcome-locked labels and 0.83 on a temporally independent external cohort of 14 antibodies. Feature attribution recovered mechanistically plausible drivers, including the diffusion interaction parameter kD and size-exclusion chromatography peak-shape metrics for viscosity and heparin-, baculovirus particle-, and poly-D-lysine- signals for exposure. These results show that routine assay panels can support practical, interpretable machine-learning guided triage of antibodies for formulation and pharmacokinetic risk, and may capture a developability-linked component of downstream progression risk.