mtDNA-FM is the first foundation model dedicated to mitochondrial DNA, a 16,569 bp circular genome. It introduces a circular positional encoding that aligns the positional period to the genome length, so the two endpoints at the D-loop junction are represented as adjacent rather than maximally distant (endpoint cosine similarity ≈0.74, versus ≈0 for standard linear encoding), together with a heteroplasmy projection channel. The 5.8M-parameter encoder is pre-trained by masked-language modeling on 152,590 vertebrate mitochondrial genomes. This preprint reports an honest zero-shot evaluation of the Phase-1 checkpoint. The model recovers some evolutionary structure without supervision but is not yet competitive: on 26-class haplogroup classification it reaches 37.9% (9.8× above chance) yet trails a supervised 6-mer baseline (78.7%), DNABERT-2 (66.3%), and even a model-free nearest-sequence lookup (79.5%); on pathogenic-variant discrimination its AUROC of 0.777 is matched by a gene-region-only baseline, indicating the signal is largely regional rather than variant-level. A per-class comparison with DNABERT-2 yields four observations that localize the shortfall to representation quality - tokenization, window aggregation, and parameter scale - and map concrete directions for a next-generation model. Code, pre-trained weights, a reproducibility pipeline, and an interactive demo are publicly available.
Thawfeek Varusai (Mon,) studied this question.