From bench assays to bedside: a context-embedding transformer predicts monoclonal antibody viscosity, clearance, and regulatory success

Late-stage failures of monoclonal antibody (mAb) programs often reflect developability liabilities, including high-concentration viscosity and rapid clearance, that standard early screens only partially capture. We introduce ACeT, an attention-based context-embedding transformer that fuses routine early-stage assay readouts to predict three endpoints: high-concentration viscosity, mouse intravenous clearance, and Phase I-to-approval outcomes. On held-out mAbs, the model achieved R2 ≈ 0.75 for viscosity and R2 ≈ 0.80 for clearance, and for clinical progression reached ~78% balanced accuracy on a 112-IgG1 cohort with outcomes locked by October 2022, then 0.83 on a temporally independent set of 14 clinical-stage mAbs whose outcomes matured by May 2025. The approach outperformed single-assay heuristics and established machine-learning baselines, and model attributions recover known biophysical drivers. Retrospective portfolio simulations, under typical industry cost and success-rate assumptions, suggest that prioritising antibodies with favourable transformer scores could reduce late-stage attrition and potentially redirect up to US$ 17 billion in capital, while also lowering animal use. By unifying heterogeneous assays in a single encoder, this framework improves the fidelity of early-stage developability decisions and offers a practical route to triage viscosity and clearance risks while prioritizing mAbs with a higher likelihood of clinical progression.