Evo 2 is an open-source artificial intelligence trained on 8.8 trillion base pairs of DNA from bacteria, archaea, and eukaryotes — essentially every major branch of life.
A DNA strand – artistic impression. Image credit: Alius Noreika / AIKey Takeaways:
- Evo 2 was trained on 8.8 trillion bases from all three domains of life plus bacterial viruses, and can identify genomic features — including splice sites, regulatory DNA, and protein-coding regions — without task-specific fine-tuning.
- The system recognized the severity of mutations, distinguished between protein-coding and non-coding RNA, and even identified which genetic code a given species uses.
- All model parameters, training code, and the OpenGenome2 dataset have been released publicly, inviting the broader research community to explore its capabilities.
Built by the same team behind the original Evo model, this 40-billion-parameter system can spot genes, regulatory sequences, splice sites, and other genomic landmarks without ever being explicitly taught what to look for. And it does this across the full spectrum of biological complexity, from the tidy genomes of bacteria to the sprawling, intron-riddled DNA of species like humans.
The original Evo, covered in late 2025, proved that feeding an AI massive amounts of bacterial genome data could produce remarkable results. Prompted with sequences from gene clusters, it could predict the next gene in line or even propose entirely novel proteins. But bacteria keep things tidy — their related genes sit side by side, neatly organized. Complex organisms? Not so much. Eukaryotic genomes are a mess by comparison: genes chopped up by introns, regulatory elements scattered across hundreds of thousands of bases, and enormous stretches of so-called junk DNA filling the gaps.
That messiness is precisely what makes eukaryotic genomes so hard to decode. The sequences that define intron boundaries or transcription factor binding sites follow loose statistical patterns rather than rigid rules. A given position might be thymine 45 percent of the time — helpful as a trend, useless as a definitive marker. Specialized bioinformatics tools exist for tasks like splice site prediction, but their error rates compound quickly across a 3-billion-base genome.
Neural networks, though, thrive on exactly these kinds of fuzzy statistical patterns. The catch has always been the sheer volume of data and compute power required. Now, with enough raw genome sequences available and a capable architecture in place, the Evo team decided to tackle the problem head-on.
How the training worked
Evo 2 runs on StripedHyena 2, a convolutional neural network. Training happened in two phases. First, the system processed sequences roughly 8,000 bases long, densely packed with important genomic features. Then it graduated to million-base-long stretches, giving it the chance to detect large-scale structural patterns.
The training data, called OpenGenome2, spans 8.8 trillion bases drawn from all three domains of life plus bacteriophages. Eukaryotic viruses were deliberately excluded — the researchers worried about potential misuse in generating sequences that could threaten humans. Two model sizes were produced: a 7-billion-parameter version trained on 2.4 trillion bases, and the full 40-billion-parameter model trained on the complete dataset.
The core principle is elegantly simple: sequences conserved across many species must be functionally important. They appear in enough contexts that the model encounters them repeatedly during training. “By learning the likelihood of sequences across vast evolutionary datasets, biological sequence models capture conserved sequence patterns that often reflect functional importance,” the researchers write. “These constraints allow the models to perform zero-shot prediction without any task-specific fine-tuning or supervision.”
That zero-shot capability matters more than it might seem. The team deliberately avoided telling Evo 2 what known splice sites look like or how regulatory elements typically behave. Fine-tuning on known features might sharpen performance on familiar patterns but could blind the system to unusual variants or entirely unknown genomic elements. Leaving it unsupervised keeps the door open for genuine discoveries.
Peeking inside the black box
To understand what Evo 2 actually learned, the researchers used interpretability tools to examine the model’s internal representations. They trained a separate neural network to decode Evo 2’s activation patterns and map them to recognizable biological features.
The results were striking. Evo 2 clearly distinguished protein-coding regions from surrounding DNA and recognized intron boundaries flanking those regions. It picked up structural features within proteins — alpha helices and beta sheets — encoded in the DNA sequence. It flagged mutations that would disrupt protein structure. It even developed internal representations for mobile genetic elements, the parasitic DNA sequences that litter most genomes.
Performance under pressure
When the team fed single-base mutations into Evo 2, the model detected problems at transcription start sites and translation initiation sites. It graded mutation severity appropriately: stop-signal mutations that would terminate protein production were flagged as more damaging than silent substitutions that left the protein intact.
The system also handled non-coding RNA with sophistication. Many cellular functions depend on RNA molecules that never get translated into protein, and Evo 2 recognized when mutations disrupted these sequences.
Perhaps most impressive, Evo 2 didn’t sacrifice its bacterial and archaeal competence to gain eukaryotic skills. It retained full performance across all domains. It even figured out which genetic code to apply — several evolutionary lineages use modified stop signals during translation, and Evo 2 correctly identified those organisms and switched to the appropriate code.
For splice site recognition, a task where eukaryotic genomes present enormous variability, Evo 2 outperformed some specialized software by certain metrics. It also showed strong results evaluating mutations in the BRCA2 gene, where many variants are linked to cancer risk. When given additional training data on known BRCA2 mutations, its accuracy improved even further.
Generating new sequences — with mixed results
The original Evo’s party trick was generating novel proteins when prompted with bacterial gene clusters. Could Evo 2 do the same with eukaryotic DNA?
The answer is complicated. When given yeast DNA as a prompt, Evo 2 produced sequences containing functional RNAs and gene-like structures complete with regulatory elements and splice sites. But the team didn’t test whether any of the predicted proteins actually worked. In bacteria, you can reasonably guess a generated gene’s function based on its neighbors. Eukaryotic genomes don’t follow that logic, making functional testing far harder to design.
A more targeted experiment asked Evo 2 to design regulatory DNA active in one cell type but not another, after providing information about sequences active in both cell types. The generated sequences were inserted into cells and tested. Only 17 percent showed activity differing by a factor of two or more between the two cell types. That’s a genuine achievement in synthetic biology — but it’s a long way from the protein-design breakthroughs some might hope for.
What comes next
Given that Evo 2 appeared less than four months after the original Evo paper, the limited experimental validation is understandable. Biology experiments take time, and choosing the right ones to run is its own challenge. Months or years may pass before the research community fully explores what this model can do for designing biologically relevant DNA.
Open questions remain about whether fine-tuned derivatives of Evo 2 could excel at specific tasks — analyzing cancer cell genomes, annotating newly sequenced species, or other specialized applications. The team appears to have prioritized rapid public release over exhaustive testing, consistent with their decision to open-source everything: model parameters, training code, inference code, and the entire OpenGenome2 dataset.
The most tantalizing possibility is one that’s hardest to test. Over the past decades, researchers have steadily discovered new types of genomic features — CRISPR repeats, microRNAs, and others that nobody knew to look for. Evo 2 may have already identified patterns corresponding to features that science hasn’t named yet. The interpretability tools described in the paper could, in theory, be used to query the model and fish out those unknown elements.
Whether anyone finds them — and what they turn out to be — is a question worth watching closely.
Reference: Nature, 2026. DOI: 10.1038/s41586-026-10176-5
Written by Alius Noreika
