Development and Experimental Validation of a Machine-Learning- and Physics- Based Exhaustive Hexapeptide Screening Model for Target Proteins.

Peptide therapeutics are rapidly advancing in the biopharmaceutical field, but the high cost and low throughput of experimental screening call for more efficient discovery pipelines. We present an AI-driven framework for exhaustive hexapeptide screening that integrates physical interaction data with high-dimensional sequence features. The two-stage pipeline first encodes peptide-protein complexes using a fragmentation-based representation to capture local pairing patterns. Then, docking-derived binding energies are combined with a transformer-based model to learn sequence-energy relationships for each target. Applied to NRP-1, STING, and cGAS, the workflow achieved high hit rates and identified novel hexapeptides with micromolar binding affinities. By bridging physical modeling with enriched feature representations, this framework offers a robust, generalizable solution that accelerates peptide drug discovery.