#2 · ML / AI methods

Structure-aware language models detect resistance genes and map variant activity

The dominant thread is the augmentation of sequence-based protein modeling with structural or dynamical information. GeoARG distills protein structural geometry into a protein language model to detect remotely homologous antibiotic resistance genes from sequence alone, which is the most distinctive methodological step among these items. Sequence Display pairs a single-round platform generating large-scale sequence-activity datasets with protein language models to map variant effects. DyME applies distributed, high-throughput MD-based mutagenesis to protein engineering and recognition mimicry, while the KRAS study uses MD-derived features with logistic regression, random forest, SVM, and Bayesian network classifiers to predict secondary-mutation drug resistance. Struct2Query converts protein pocket similarity into composite-molecule ROCS queries for structure-guided virtual screening. What is new is the use of geometric distillation to transfer structural knowledge into sequence-only models.