Fig. 7: Scorpio Models Reveal AMR Prediction Strength and Functional Insights.
From: Enhancing nucleotide sequence representations in genomic analysis with contrastive optimization
a Comparison of Antimicrobial Resistance (AMR) prediction performance metrics across different models. This table highlights that Scorpio models outperform other models in AMR tasks, even though they were not explicitly trained on the AMR dataset, using gene-taxa-based training instead. (Bold font is used for the best results, and underline is used for the second-best results). b Cross-attention analysis of two AMR genes (KamB and Kmr) conducted using the Scorpio-BigDynamic model. High-attention regions identified by the model include critical areas such as the βN1 and βN2 (orange) regions, conserved regions (blue and green), mutation sites (purple), and the β6/7 linker (yellow). These regions are predominantly located at the junctions of α-helices and β-sheets, suggesting functional relevance as detected by the model. c Sequence logo of the KamB and Kmr genes, with letters highlighted based on the regions with high attention identified by the model.
