OPEN LINK ↗
// 71d agoBENCHMARK RESULT
Evo 2 embeddings surface promoter links BLAST misses
A Reddit experiment using Evo 2 intermediate embeddings across 25 human genes found at least one strong biologically plausible promoter-level match (VIM vs DES) that standard BLAST alignment could not detect. The author’s heavy filtering removed many repeat-driven false positives, suggesting real regulatory signal may be present but still hard to extract reliably.
// ANALYSIS
This is an intriguing early sign that genomic foundation models can encode functional biology beyond sequence identity, but the signal-to-noise ratio is still the main bottleneck.
- –The VIM/DES hit is compelling because it links co-regulated, related genes despite no detectable sequence alignment.
- –Most top matches were driven by repetitive elements (especially Alu), showing how easily embeddings can latch onto confounders.
- –The need for strict post-filtering suggests current workflows are research-grade, not yet robust enough for routine discovery pipelines.
- –If replicated at scale, this kind of embedding-space retrieval could complement alignment-based methods for non-obvious regulatory hypotheses.
// TAGS
evo-2llmembeddingbenchmarkresearchopen-source
DISCOVERED
71d ago
2026-03-17
PUBLISHED
72d ago
2026-03-17
RELEVANCE
7/ 10
AUTHOR
Clear-Dimension-6890