Rapid in-silico identification of Crabtree-positive and Crabtree-negative yeasts from mitochondrial genomes is essential for guiding experimental design when annotations are missing or conserved motifs are unknown. In this study, we analyzed mitochondrial sequences from 64 yeast species (34 Crabtree-positive and 30 Crabtree-negative) and obtained new, informative representations of these DNA sequences by probing and fine-tuning the genomic foundation model HyenaDNA. To address the limitations of conventional aggregation methods such as max or mean pooling, which risk loss of contextual information, and to better capture long-range dependencies inherent in genomic sequences, we introduced a novel similarity-based embedding aggregation method, “Sim Pooling.” Using this approach, fixed-length embeddings were generated and subsequently applied in downstream prediction tasks. Five classical machine learning classifiers—logistic regression, support vector classifier (SVC), random forest, eXtreme Gradient Boosting (XGBoost), and K-Nearest Neighbor—were trained and evaluated using metrics including accuracy, F1-score, AUROC, specificity, and sensitivity. Logistic regression trained on fine-tuned HyenaDNA embeddings with Sim Pooling achieved the best performance, yielding a mean F1-score of 0.79 ± 0.04 across 5-fold cross-validation, outperforming both traditional k-mer – based and probing-based models. These results demonstrate that fine-tuned genomic language models, combined with innovative embedding aggregation strategies such as Sim Pooling, can reliably predict the metabolic fate of yeast strains from raw mitochondrial sequences. With modest refinements and expansion to larger datasets, this framework can be extended to identify motifs or features driving metabolic states, reducing dependence on wet-lab validation or classical bioinformatics screening.