Transformer-Guided Co-Optimization of Mutation Prioritization and Thermodynamic Probe Design for Mutation Detection

The design of sequence-selective DNA probes has long been limited by a decoupled workflow in which target selection proceeds without considering the physicochemical determinants of probe performance, such as thermodynamic stability and mismatch discrimination. Here, we establish an iterative computational design strategy that unifies mutation prioritization with chemical probe design. A custom transformer model extracts diagnostic attribution scores of mutations from 1,782 samples, generating semantic maps to guide probe placement. Candidate designs undergo in silico thermodynamic screening based on differential binding free energy. Probes failing to satisfy the integrated criteria of thermodynamic discrimination, diagnostic attribution, and population frequency are further evaluated by refining their spatial configuration to co-optimize thermodynamic discriminability and clinical relevance. This iterative process converges on an 8-plex targeted panel. Wet laboratory validation confirms 0.03% variant allele frequency detection with low assay cost and short turnaround time. Clinical testing on 375 samples achieves AUROC values of 0.96 for tissue and 0.90 for plasma. This work provides a potential chemical framework for nucleic acid probe design that could be explored in other molecular diagnostic systems.