Deep Learning-Enabled Engineering of a Hyper-Stable and Soluble MPB70-83 Antigen for Sensitive Bovine Tuberculosis Surveillance.

Abstract

Bovine tuberculosis (bTB) diagnostics are frequently hindered by the structural instability and insolubility of recombinant multi-epitope proteins in prokaryotic expression systems. To address this, we developed an integrated strategy combining immunoinformatics with the ProteinMPNN deep learning framework. We constructed a fusion protein utilizing B-cell epitopes from MPB70 and MPB83, where ProteinMPNN was employed to redesign non-epitope scaffolds for optimized thermodynamic stability. The structural integrity was verified via 200-ns molecular dynamics (MD) simulations. MD analysis revealed the redesigned construct transitions into a rigid, compact native state with a stable hydrophobic core. Validating this in silico design, the optimized protein achieved high-yield soluble expression in Escherichia coli, eliminating inclusion body formation. The resulting indirect ELISA demonstrated 96.30% sensitivity and 98.61% specificity, showing excellent concordance (Cohen's κ > 0.9) with standard assays. This study demonstrates that deep learning-based sequence redesign effectively resolves solubility bottlenecks in antigen engineering, providing a robust, scalable tool for precision bTB surveillance.