Fig. 1: Overview of the Scorpio framework.

A Gene-Taxa Dataset Creation: genomes from NCBI were downloaded using the Woltka pipeline²⁷ and filtered to include 497 named genes from 1929 genera (a single-species representative per genus). This process removed most unknown and hypothetical proteins and focused on the most common, conserved, and well-studied genes, particularly housekeeping genes. Results of filtering are shown as a barplot, and the distribution of samples per level is shown in a box plot, indicating a balanced dataset at the gene level. B Training and Inferring with Scorpio: DNA sequences are encoded using 6-mer frequency and BigBird embeddings. The configuration supports different Scorpio models, such as Scorpio-6Freq, Scorpio-BigDynamic, and Scorpio-BigEmbed, with adjustable hierarchical levels for enhanced generalization, allowing adaptation to different datasets and hierarchies. During inference, one triplet branch is used to obtain the embedding vector, which is the final layer of the network. C Indexing and Searching: FAISS is utilized for efficient embedding retrieval of each query and to find the nearest neighbor. Based on the nearest neighbor from the validation set, we train a confidence score model at each level of the hierarchy. During inference, this model calculates the confidence for each query. Depending on the application, classification results and confidence scores are reported.