Rational Inhibitor Discovery for BRAF^V600E Using PSeMut, a Sequence-Driven Model

Mutational heterogeneity at drug targets both undermines small-molecule efficacy and contributes to disease. A practical response is to discover scaffolds that maintain potency in the presence of mutations; however, prevailing biological research strategies are not readily scalable, constraining high-throughput design and discovery. We present PSeMut, a structure-free Siamese model that contrasts wild-type and mutant protein-ligand fingerprints (PSICHIC-derived) to predict mutation-induced activity changes. On a variant-resolved benchmark, PSeMut attains a test RMSE of 0.400 ± 0.025 and consistently outperforms classical baselines trained on identical features and splits; removing the exchange-consistency constraint degrades performance. To evaluate the feasibility of the framework in a prospective setting, we applied a structure-free prioritization pipeline-combining scaffold-novelty filtering, PSICHIC activity scoring, PSeMut-based mutation-tolerance ranking, clustering-based selection, and biological validation. This workflow prioritized SNS-314 for follow-up testing. In the selected validation assays, SNS-314 showed mutation-selective cellular activity (IC50 = 0.45 μM in BRAF^V600E; 7.5-fold vs. BRAF^WT), suppressed the BRAF-MEK-ERK signaling axis, and produced about 50% tumor growth inhibition in vivo without overt toxicity. Together, these results validate the effectiveness of PSeMut in a structure-free, mutation-aware screening workflow that links sequence-driven modeling to experimental confirmation and enables rational prioritization of mutation-resilient scaffolds across heterogeneous disease settings.