The Virtual Cell Challenge's final test set is now available at virtualcellchallenge.org. Final scores will be based on the 100 new perturbations. The 7 day phase is single blind (no score feedback) and uses the latest cell-eval (bugfix'd last Friday). Good luck everyone!
Zach believes we're actually making a company to sell this thing. He has a business card (he's the CEO and I'm his CTO obvs). Even made a badge. For now, we're open sourcing the base model :). Python code and build instructions here: github.com/daveyburke/Z.... Enjoy!
My 7 yr old has been dreaming up an AI teddy bear. So we made him! Zaby is a clever, pedagogical & funny teddy that loves talking math. Powered by Gemini Flash & Google speech recognition/synthesis. His mouth moves in sync with the speech envelope. Smarter than the average bear!
"They help us unravel... the language of life" - Arc Institute's Evo 2 model featured during NVIDIA's GTC 2025 keynote - m.youtube.com/watch?v=_waP...
Excited to announce the Arc Virtual Cell Atlas, a collection of high quality, curated, open datasets, incorporating scBaseCamp and Tahoe-100M from @vevo_ai. We hope this can be the beginning of an "ImageNet moment" for virtual cell modeling. Available at arcinstitute.org/tools/virtua...
In the future, we can use this mechanism to steer DNA generation, for example make a prokaryotic sequence have more eukaryotic features, or increase the presence of alpha helices. You can read more in the Evo 2 preprint here: arcinstitute.org/manuscripts/...
It shows genomic concepts in a reference genome such as coding sequences, alpha helices, tRNAs, etc. The tool overlays corresponding features that activate when Evo 2 detects such concepts. What’s amazing is Evo learned all this from genomes in nature without any supervision!
Together with @GoodfireAI we built a visualizer that lets you explore the concepts learned by Evo 2. Try it here: arcinstitute.org/tools/evo/ev...
We applied sparse autoencoders to Evo 2, our new DNA model, to show it autonomously learns a breadth of biological features, including exon–intron boundaries, transcription factor binding sites, protein structural elements, and prophage genomic regions
A common critique of large AI models is that they are black boxes. The recent field of mechanistic interpretability aims to “look inside” the AI black box
This is one of many applications of this work. Evolution has learned to read and write DNA over millions of years, and Evo 2 aims to learn from this knowledge. The AI model serves as a foundation for understanding the language of life across all domains—from bacteria to humans
This particular variant was initially reported as a variant of unknown significance (VUS). Years later, oncologists learned it was a driver of breast and ovarian cancers. In the Evo paper, we show state of the art performance on classifying BRCA1 variants of unknown significance
If I take a known deleterious mutation c.5095C>T that changes just the 5095th nucleotide in exon 17 from C to T, the negative log likelihood increases from 0.96 to 0.99 indicating the model is less confident. Evo recognizes that this mutation causes a loss of function of the gene
Evo Designer can also score DNA sequences, i.e. how likely the sequence is in nature. Here’s an example of a section of the BRCA1 - certain mutations in this gene are known to increase the risk of breast & ovarian cancer
Prompt with a sequence or species and the model will generate new DNA. Select sections of generated DNA to visualize the corresponding proteins, or use BLAST to find similar sequences in nature
We built a new interactive user interface for generation and scoring called Evo Designer arcinstitute.org/tools/evo/ev...
Evo 2 uses a new hybrid architecture called StripedHyena 2 enabling a long context window of 1M nucleotides with a model size of 40B parameters, trained on 2048 H100 GPUs. Preprint can be found at arcinstitute.org/manuscripts/... and includes links to source code
Introducing Evo 2 from Arc Institute - an AI that can model and design the genetic code for all domains of life. It’s one of the largest-scale truly open source AI models for biology (and in fact more generally - most “open source” large language models are only “open weights”)
