Evo 2 genomic language model fine-tuned on 750M prokaryotic intergenic sequences to cluster noncoding RNAs that sequence alignment can't relate.
No providers recorded yet. Browse all providers
Two riboswitches that bind the same metabolite can fold into the same three-dimensional pocket while sharing barely a quarter of their DNA sequence. Alignment-based tools, which relate sequences by counting matching bases, lose track of such pairs in the "twilight zone" below roughly 50–60% identity, so most of the noncoding DNA between prokaryotic genes has never been sorted into families. Evo-PIN (Prokaryotic INtergenic) addresses this by letting a genomic language model judge relatedness instead: sequences that the model represents similarly are grouped together, whether or not they align.
Evo-PIN was developed at the Arc Institute in the Konermann and Hsu labs, with co-authors at Stanford University and Heidelberg University, and described in a bioRxiv preprint posted in October 2026. It is the 7B-parameter, 1M-context Evo 2 model after continued pretraining on prokaryotic intergenic DNA. Intergenic regions make up only about 10% of prokaryotic genomes, so they were thinly represented in Evo 2's original training data. Further next-nucleotide training on them sharpened the model's representation of the covarying base pairs that hold structured RNAs together.
Evo-PIN drives COMPASS (COMPrehensive Atlas of Structured Sequences), a pipeline and database that clusters intergenic sequences using frozen Evo-PIN embeddings. Every readout in the study comes from this single fixed checkpoint, without task-specific supervision.
The training corpus combined 465 Gbp from 143,614 representative bacterial and archaeal species in GTDB release 226 with 855 Gbp of metagenomic sequence from OpenGenome2. Gene boundaries called by Prodigal, plus tRNA and rRNA annotations, were used to cut out 1.01 billion intergenic regions. MMseqs2 deduplication at 70% identity reduced these to 750 million representative sequences. All parameters of Evo 2 7B, a StripedHyena 2 hybrid of convolutional and attention layers, were then fine-tuned on these sequences and their reverse complements with a causal language-modeling objective at 2,048-token context. Checkpoints were scored by how well their embeddings clustered sequences sharing an Rfam annotation, and the 400K-step checkpoint, reached after about 35% of one epoch on 32 H100 GPUs, was named Evo-PIN.
On a 10-million-sequence benchmark covering 466 ncRNA families, the normalized cluster-quality score rose from 0.43 for mean-pooled Evo 2 embeddings to 0.60 for Evo-PIN, and to 0.65 with low-entropy pooling. Embeddings are taken from the 4,096-dimensional pre-normalization output of block 25. Median Evo-PIN entropy inside annotated ncRNAs was 73% lower than outside them, against 55% for Evo 2. An approximate k-nearest-neighbor graph with Leiden clustering grouped 376 million intergenic sequences into 5.14 million clusters. Among the 24,837 clusters with at least five annotated members, annotation consistency averaged 0.92, while median pairwise sequence identity within clusters was 28.5%.
Evo-PIN is aimed at microbiologists and RNA biologists mining genomes and metagenomes for uncharacterized noncoding elements. Through COMPASS, the authors nominated 20,149 clusters of putative structured RNAs that have a candidate protein partner, and found that fewer than 1% of all clusters carried a known ncRNA annotation. Targeted reclustering of sequences adjacent to bridge recombinases identified bridge RNA variants with longer target-recognition sequences than any characterized ortholog, which is of direct interest for genome engineering. Under researcher direction, LLM agents querying the MCP server surfaced divergent riboswitches and hammerhead ribozymes that fall below their own covariance-model thresholds, along with long candidate ncRNAs encoded next to PhoH2 proteins. All of these candidates are computational predictions awaiting experimental testing.
Evo-PIN shows that a short round of domain-focused continued pretraining can turn a general genomic language model into a reusable instrument for discovery, with four separate readouts (embeddings, entropies, entropy periodicity and categorical Jacobians) coming from one frozen checkpoint. The preprint has not yet been peer reviewed. The COMPASS atlas is openly released under CC BY 4.0, but the Evo-PIN weights and the fine-tuning code have not been released, so the model itself can only be reached through its precomputed outputs.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.