Feedback from Evo

Artificial intelligence foundation models that treat DNA as a language open new avenues for sequence prediction and design. Evo (2024) is among the first long-context genomic LLMs operating at single-nucleotide resolution, enabling next-token generation conditioned on large genomic prompts. We benchmarked Evo for microbial sequence continuation to assess fidelity, scaling behavior, and practical usability for genomics/pangenomics. Using a diverse panel of microbial genomes, we sampled random genomic loci to create prompts spanning 1–1000 nt, paired with target continuation lengths of 1–1000 nt. For each (prompt, length) combination, Evo generations were evaluated by global (Needleman–Wunsch with affine gaps) and local (Smith–Waterman) alignments against the true genomic continuation, reporting percent identity and query coverage; negative controls used minimal prompts.Empirically, alignment identity degrades with longer generations and with shorter prompts, consistent with cumulative next-token errors and weakened long-range conditioning. While occasional local matches suggest recovery of accurate motifs, Evo's outputs are insufficiently faithful for reliable microbial genome synthesis.However, we also show that this approach holds promise and outline immediate extensions: benchmarking Evo 1.5, Evo2 (species-guided conditioning, larger context window), probing temperature/top-k/top-p effects, conditioning by taxon or region (core vs. accessory), BLAST-based functional concordance, and targeted fine-tuning.