AI Science Tracker

Arc Institute

6 discoveries from 2024 to 2026

Discoveries

Evo 2 uses in-context learning to design de novo regulatory sequences with only tens of examples

Arc Institute reports that Evo 2 designs functional regulatory sequences from 4 to 64 in-context examples without extra training, with 30-50% yield. Of nine AI-designed enhancers delivered by AAV to mice, seven drove detectable expression and two labeled layer 5 extratelencephalic neurons with 73% and 95% specificity. Results are bioRxiv preprints; in vivo work covers only layer 5 enhancers, and other designs (proteins, CRISPR loci, antibodies, phage genomes) are computational only.

Technology: Evo 2 (genomic language model); in-context learning; reinforcement learning post-training

Family Evo: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Nine AI-designed enhancers were delivered by AAV to mice: seven drove detectable expression and two labeled layer 5 extratelencephalic neurons with 73% and 95% specificity.
  • Still unknown: Only layer 5 enhancers were tested in vivo; the authors plan other cell types, and designs beyond enhancers (proteins, CRISPR loci, antibodies, phage genomes) are computational only.
  • AI role: Evo 2 generated regulatory sequences from 4 to 64 in-context examples; reinforcement learning with predictive reward models sharpened cell-type specificity.

Evidence sources: https://arcinstitute.org/news/evo-2-in-context-learning

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: not linked in this catalog; this does not mean no paper exists.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Bolero predicts cell-state-specific gene regulation from DNA across species

Bolero predicted cell-specific responses to AP-1 activity and transferred a human-trained response model to mouse and macaque neuronal enhancers, according to Arc Institute. Its announcement also describes a computational AP-1/cofactor spacing hypothesis; these virtual edits do not establish biological causality. This summary is limited to the October 20, 2026 institutional announcement, which identifies a collaboration with Harvard Medical School and describes a 10.8-million-cell atlas. The linked paper is a preprint; full-paper methods and experimental independence were not assessed.

Technology: Bolero (cell-state-conditioned sequence-to-function Transformer)

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: Human-trained AP-1 response predictions compared with mouse/macaque neuronal enhancers; virtual regulatory perturbations.
  • Still unknown: Source checking does not establish independent confirmation. Review date: 2026-11-15.
  • AI role: Bolero, a Transformer conditioned on cell-state embeddings, predicts cell-state-specific gene regulation directly from DNA and estimates responses in unprofiled cell states.

Evidence sources: https://arcinstitute.org/news/bolero

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Evo genomic language model designs functional toxin-antitoxin and anti-CRISPR systems

Evo used genomic context to generate functional bacterial defence-system components. The paper reports an AI-generated toxin and low-sequence-similarity antitoxins that neutralized it in growth assays, including antitoxins that also acted against multiple natural toxins; it also reports five generated anti-CRISPR proteins that protected bacteria against Cas9 activity in cell and phage assays. Two generated anti-CRISPR proteins lacked significant sequence similarity to known proteins. These experiments demonstrate function-guided genomic design beyond simple copying of natural sequences.

Technology: Evo 1.5, a long-context generative genomic language model trained on prokaryotic DNA sequences

Family Evo: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: The paper reports an AI-generated toxin and low-sequence-similarity antitoxins that neutralized it in growth assays, including antitoxins that also acted against multiple natural toxins; it also reports five generated anti-CRISPR proteins that protected bacteria against Cas9 activity in cell and phage assays.
  • Still unknown: This is a set of functional in-vitro bacterial assays, not evidence of therapeutic use, broad host range or ecological performance. One generated toxin had 71% sequence identity to a known RelE toxin; novelty is especially supported by low-similarity antitoxins and anti-CRISPRs, two of which lacked significant sequence identity to known proteins. Earlier Evo phage-genome design and Evo 2 regulatory-sequence work are different papers and different functional results; this row is limited to the toxin-antitoxin/anti-CRISPR findings in the cited Nature article.
  • AI role: Researchers prompted Evo 1.5 with genomic contexts that associate functionally related bacterial genes and used it to generate candidate toxin-antitoxin and anti-CRISPR sequences. Candidates were filtered computationally, synthesized/expressed and tested in bacterial growth, Cas9-protection and phage assays. AI generated candidates; the biological activity was established experimentally.

Evidence sources: https://www.nature.com/articles/s41586-025-09749-7.pdf

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Arc Institute

First bacteriophage genomes designed by a genome language model

The Arc Institute group used Evo to design genomes resembling that of phage ΦX174: of 285 designs, 16 yielded functional phages (67–392 mutations from the closest natural one; 13 had never-seen mutations). They infect only two E. coli strains; a cocktail overcame resistance in three evolved strains. bioRxiv preprint; peer review is unconfirmed.

Technology: Evo (genome language model)

Family Evo: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: direct test
  • Tested: Of 285 Evo-designed genomes, 16 were synthesized and gave functional phages in E. coli, and these were tested for fitness, host range and, for three resistant strains, a phage cocktail.
  • Still unknown: The source states host range was limited to E. coli C and W, all 16 phages were based on the phiX174 template, and clinical use is not shown.
  • AI role: Evo, fine-tuned on Microviridae sequences, generated whole phage genomes resembling phiX174.

Evidence sources: https://arcinstitute.org/news/hie-king-first-synthetic-phage · https://www.biorxiv.org/content/10.1101/2025.09.12.675911

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 5/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Arc Institute

Evo 2: the largest AI model for biology

Trained on more than 9.3 trillion nucleotides from over 128,000 genomes, it reads sequences up to 1 million nucleotides, flags harmful mutations, and designs genetic elements and genomes as long as those of simple bacteria. In BRCA1 tests it exceeds 90% accuracy in distinguishing benign from potentially pathogenic variants (a result limited to that gene). Preprint; human pathogens were excluded from training.

Technology: Evo 2 (genome language model)

Family Evo: profile · models page

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Preprint
  • Test type: retrospective or benchmark
  • Tested: The model was tested on BRCA1 variants, where it reached over 90% accuracy distinguishing benign from potentially pathogenic mutations, and on identifying genetic changes affecting protein function and fitness.
  • Still unknown: The source reports the BRCA1 result for one gene only and states no other limitations beyond human pathogens being excluded from training.
  • AI role: A genome language model trained on 9.3 trillion nucleotides flags harmful mutations and designs genetic elements.

Evidence sources: https://arcinstitute.org/news/evo2

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Editorial assessment: Impact 4/5 · Validation 2/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Evo: a foundation model for DNA, from molecules to genomes

Published in Science, Evo (Arc Institute, Stanford, and UC Berkeley) designed EvoCas9-1, a CRISPR system sharing only about 73% of its sequence with CRISPR-Cas9 but with comparable activity after eleven designs, and new IS200/IS605 transposons that cut and paste DNA. The team calls this the beginning and conducted an ethics and safety study.

Technology: Evo (DNA foundation model)

Family Evo: profile · models page

Publication: Peer-reviewed paper · Test context: direct test. Catalog categories, not an evidence grade.

Result, AI contribution and limits

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

  • Publication status: Peer-reviewed paper
  • Test type: direct test
  • Tested: Evo was prompted to generate CRISPR systems; EvoCas9-1 was obtained after testing eleven designs, validated in the lab by an experimental biology subteam.
  • Still unknown: The page gives no success rates or error analysis, and an ethics and safety investigation addressed potential risks rather than efficacy.
  • AI role: Evo, a DNA-trained foundation model, designed the protein and RNA components of EvoCas9-1 and new IS200/IS605 transposons.

Evidence sources: https://arcinstitute.org/news/evo-science

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: press release. Paper: linked separately.

Review scope and date are not recorded; source checked does not establish full-text review.

Citations: in the top 1% of works of the same type, year and subfield (OpenAlex work metrics, 737 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.

Editorial assessment: Impact 4/5 · Validation 3/5 · Novelty 4/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Follow-up and relationships

No follow-up recorded; this is not a claim that no follow-up exists.

Follow-up review scope and date are not recorded.

No source-supported relationships recorded.

Open the interactive profile for the timeline, or browse all organizations and model families.