In Silico Directed Evolution of Humanized Peptide Transporters via Attention-Guided Epistatic Rescue

The human intestinal peptide transporter 1 (SLC15A1, PepT1) is a critical determinant of oral drug bioavailability, yet its thermodynamic characterization and structural tractability remain challenging. While bacterial orthologues like the Escherichia coli DtpA transporter offer highly stable structural surrogates, significant cross-species variations in the binding microenvironment limit their pharmacological fidelity. Traditional structure-based engineering aimed at engrafting the human pharmacophore onto bacterial scaffolds is frequently hindered by combinatorial explosion and severe thermodynamic frustration, as non-native side-chains disrupt co-evolved local packing. To bypass the empirical bottlenecks of in vitro directed evolution, we present a fully in silico co-evolutionary pipeline leveraging deep contextual protein language models (PLMs). By mapping 19 pharmacologically critical human PepT1 residues onto the DtpA sequence, we utilized the 650-million parameter ESM-2 model to conduct zero-shot mutational profiling. We identified 14 primary humanizing mutations that triggered profound local destabilization (fitness probabilities < 10 − 5 ). To resolve these energetic conflicts, we deployed an automated, attention-guided epistatic rescue algorithm to sweep the flanking microenvironmental topologies. This computationally efficient heuristic successfully identified localized, non-native compensatory mutations for all high-risk targets without relying on traditional molecular dynamics. Most notably, the highly deleterious Q41R and R305F substitutions were buffered by Y38D and I304K secondary mutations, yielding 5,215-fold and 11,614-fold improvements in contextual stability, respectively. Binding dynamics is confirmed via docking calculations and all-atom membranous molecular dynamics simulations. This deep learning-driven framework successfully rationalizes the engraftment of the human binding pocket, drastically reducing the required experimental screening space and yielding a thermodynamically robust, humanized surrogate primed for advanced drug discovery.