Sequence modeling and design from molecular to genome scale with Evo

May combine two claims.

Evo was trained on OpenGenome, a dataset containing more than 80,000 bacterial and archaeal genomes plus millions of predicted phage and plasmid sequences, totaling 300 billion nucleotide tokens.

“We compiled a large genome dataset called OpenGenome (Materials and methods) with more than 80,000 bacterial and archaeal genomes and millions of predicted phage and plasmid sequences, covering 300 billion nucleotide tokens”

• cites [35] Parks DH, GTDB: An ongoing census of bacterial and archaeal diversity through a phylogenetically consistent, rank normalized and complete genome-based taxonomy. Nucleic Acids Res. 50, D785– D794 (2022). doi: 10.1093/nar/gkab776 34520557 PMC8728215 sourceCould not verify

  The Introduction and Data Availability sections describe GTDB as a bacterial and archaeal genome taxonomy, not OpenGenome or Evo’s training data. Searches for OpenGenome, 80,000 genomes, phage and plasmid sequences, and 300 billion nucleotide tokens found no passage establishing the claimed dataset composition or size. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [36] Camargo AP, IMG/VR v4: An expanded database of uncultivated virus genomes within a framework of extensive functional, taxonomic, and ecological metadata. Nucleic Acids Res. 51, D733– D743 (2023). doi: 10.1093/nar/gkac1037 36399502 PMC9825611 sourcePartly supported

  The 'Automatic virus identification from IMG datasets' section (p. 3) says the authors processed 83,858 bacterial and archaeal isolate genomes, and the Abstract (p. 2) reports more than 15 million viral genomes and fragments. The paper does not establish that those sequences formed OpenGenome or were used to train Evo, nor does it give a 300-billion-token total. The Conclusion (p. 11) describes a plasmid-focused database as still in development, rather than reporting millions of plasmid sequences in this collection. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [37] Camargo AP, IMG/PR: A database of plasmids from genomes and metagenomes with rich annotations and metadata. Nucleic Acids Res. 52, D164– D173 (2024). doi: 10.1093/nar/gkad964 37930866 PMC10767988 sourceNot checked — no full text of the source

May combine two claims.

Existing DNA language models are limited by dense Transformer architectures, incur high computational cost that scales quadratically with sequence length, and generally underperform at single-nucleotide or byte-level resolution compared with models trained at coarser resolutions.

“existing attempts to model DNA as a language (15–17) are limited by the prevailing dense Transformer architecture, which incurs high computational cost as input sequence lengths grow relative to model width (scaling quadratically) and generally underperforms at single-nucleotide or byte-level resolution compared with models trained at coarser resolutions (18).”

• cites [15] Zvyagin M., GenSLMs: Genome-scale language models reveal SARS-CoV-2 evolutionary dynamics. Int. J. High Perform. Comput. Appl 37, 683– 705 (2023). doi: 10.1177/10943420231201154 sourcePartly supported

  Split into 2 atoms — 0 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [16] Dalla-Torre H., The Nucleotide Transformer: Building and Evaluating Robust Foundation Models for Human Genomics. bioRxiv 2023.01.11.523679 [Preprint] (2023); 10.1101/2023.01.11.523679. PMC11810778 39609566 sourcePartly supported

  Split into 2 atoms — 1 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [17] Zhou Z, Ji Y, Li W, Dutta P, Davuluri R, Liu H, DNABERT-2: Efficient foundation model and benchmark for multi-species genome. arXiv:2306.15006 [q-bio.GN] (2023). sourcePartly supported

  Split into 2 atoms — 1 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [18] Tay Y., Charformer: Fast Character Transformers via Gradient-based Subword Tokenization. arXiv: 2106. 12672 [cs.CL] (2022). sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

For some genomes, Evo zero-shot essentiality prediction with 66k context is strong, with AUROC 0.90 on lambda phage and 0.84 on Pseudomonas aeruginosa.

“For a few genomes, the zero-shot performance with 66k context is notably strong, with an AUROC of 0.90 on lambda phage essentiality data (74) and an AUROC of 0.84 on Pseudomonas aeruginosa essentiality data (75)”

• cites [74] Piya D., Systematic and scalable genome-wide essentiality mapping to identify nonessential genes in phages. PLOS Biol. 21, e3002416 (2023). doi: 10.1371/journal.pbio.3002416 PMC10695390 38048319 sourceSupported
• cites [75] Turner KH, Wessel AK, Palmer GC, Murray JL, Whiteley M, Essential genome of Pseudomonas aeruginosa in cystic fibrosis sputum. Proc. Natl. Acad. Sci. U.S.A 112, 4110– 4115 (2015). doi: 10.1073/pnas.1419677112 25775563 PMC4386324 sourceNot checked — no full text of the source

May combine two claims.

Evo's zero-shot correlation on promoter-RBS sequences exceeded GC content (r = 0.47), GenSLM (r = 0.11), and RBS Calculator (r = 0.39).

“Evo’s zero-shot correlation on promoter-RBS sequences is also higher than the GC content of the promoter-RBS sequences (Spearman r = 0.47), zero-shot GenSLM likelihoods (Spearman r = 0.11), and RBS Calculator (Spearman r = 0.39)”

• cites [57] Salis HM, Mirsky EA, Voigt CA, Automated design of synthetic ribosome binding sites to control protein expression. Nat. Biotechnol 27, 946– 950 (2009). doi: 10.1038/nbt.1568 19801975 PMC2782888 sourceSupported
• cites [58] Reis AC, Salis HM, An automated model test system for systematic development and improvement of gene expression models. ACS Synth. Biol 9, 3145– 3156 (2020). doi: 10.1021/acssynbio.0c00394 33054181 sourcePartly supported

  The section “A systematic comparison of translation initiation rate models” (page unavailable) identifies RBS Calculator as a model for predicting translation initiation rates. It does not report a Spearman correlation of 0.39 on promoter-RBS sequences or compare the model with Evo, GC content, or GenSLM; those performance claims are not supported by this source. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

Transformer-based DNA models are constrained to short context lengths and use token aggregation schemes, sacrificing single-nucleotide resolution.

“As a result, Transformer-based DNA models are constrained to short context lengths and use schemes that aggregate nucleotides into the basic units of language models, called tokens, thereby sacrificing single-nucleotide resolution (15, 16, 21–23).”

• cites [15] Zvyagin M., GenSLMs: Genome-scale language models reveal SARS-CoV-2 evolutionary dynamics. Int. J. High Perform. Comput. Appl 37, 683– 705 (2023). doi: 10.1177/10943420231201154 sourcePartly supported

  Section 4.2 contrasts standard 1,024- or 2,048-token contexts with the longer sequences needed for genomes, and Section 5.2.1 says GenSLM groups three nucleotides into each codon-level input token. But the paper does not say this sacrifices the ability to model individual nucleotide changes; Section 5.2.1 says GenSLM models individual mutations at nucleotide scale. The claim therefore overstates the resolution consequence and treats a common context-length challenge as a constraint on Transformer-based DNA models generally. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [16] Dalla-Torre H., The Nucleotide Transformer: Building and Evaluating Robust Foundation Models for Human Genomics. bioRxiv 2023.01.11.523679 [Preprint] (2023); 10.1101/2023.01.11.523679. PMC11810778 39609566 sourceSupported
• cites [21] Fishman V., GENA-LM: A family of open-source foundational DNA language models for long sequences. bioRxiv 2023.06.12.544594 [Preprint] (2024); 10.1101/2023.06.12.544594. PMC11734698 39817513 sourceSupported
• cites [22] Ji Y, Zhou Z, Liu H, Davuluri RV, DNABERT: Pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome. Bioinformatics 37, 2112– 2120 (2021). doi: 10.1093/bioinformatics/btab083 33538820 PMC11025658 sourcePartly supported

  Methods §2.2 says DNABERT uses k-mer tokens rather than individual bases and accepts at most 512 tokens per input, supporting the tokenization and context-limit portions of the claim. But the same section describes a method for processing longer sequences, and the Results state that DNABERT enables visualization of nucleotide-level importance. The paper therefore does not support the unqualified claim that token aggregation sacrifices single-nucleotide resolution. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [23] Hwang Y, Cornman AL, Kellogg EH, Ovchinnikov S, Girguis PR, Genomic language model predicts protein co-regulation and function. Nat. Commun 15, 2880 (2024). doi: 10.1038/s41467-024-46947-9 38570504 PMC10991518 sourcePartly supported

  The Introduction (page unavailable) describes GenSLMs as using codon-level representations and says that learning gene-to-gene interactions requires longer genomic segments. That supports an example of nucleotide aggregation and a need for longer context, but it does not establish that Transformer-based DNA models generally have short context lengths or sacrifice single-nucleotide resolution. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The DNA-targeting Cas9 nuclease is typically encoded within 3000 to 4800 base pairs and is found in close genomic proximity to its cognate CRISPR array.

“The DNA-targeting Cas9 nuclease is typically encoded within 3000 to 4800 base pairs (bp) of coding sequence and found in close genomic proximity to its cognate CRISPR array (60).”

• cites [60] Hsu PD, Lander ES, Zhang F, Development and applications of CRISPR-Cas9 for genome engineering. Cell 157, 1262– 1278 (2014). doi: 10.1016/j.cell.2014.05.010 24906146 PMC4343198 sourceSupported

Classified as probably not a substantial claim.

A single-nucleotide mutation in an essential gene can be incompatible with life if it disrupts that gene's expression or function.

“even a single-nucleotide mutation in an essential gene can be incompatible with life if it disrupts that gene’s expression or function (72).”

• cites [72] Rocha EPC, Danchin A, Gene essentiality determines chromosome organisation in bacteria. Nucleic Acids Res. 31, 6570– 6577 (2003). doi: 10.1093/nar/gkg859 14602916 PMC275555 sourcePartly supported

  The Introduction describes essential genes as serving indispensable cellular functions and discusses identifying them through gene inactivation. The ‘Essentiality drives gene strand bias’ section says nonfunctional peptides are especially counter-selected in essential functions. Neither section addresses whether a single-nucleotide mutation disrupting an essential gene’s expression or function can be incompatible with life. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Diverse families of CRISPR-Cas systems are found throughout bacterial and archaeal life, including Cas12- and Cas13-based systems that target DNA and RNA, respectively.

“Diverse families of CRISPR-Cas systems are found throughout bacterial and archaeal life, such as Cas12- or Cas13-based systems that target DNA and RNA, respectively (61).”

• cites [61] Koonin EV, Makarova KS, Origins and evolution of CRISPR-Cas systems. Phil. Trans. R. Soc. B 374, 20180087 (2019). doi: 10.1098/rstb.2018.0087 PMC6452270 30905284 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Sampling directly from a large natural language model usually yields text that is grammatical but locally favors simpler sentences and is globally incoherent, particularly over long spans.

“For example, directly sampling from a large natural language model typically produces sequences that are grammatically correct yet locally biased toward simpler sentence constructions and that are globally incoherent, especially at long lengths.”

• cites [83] Wei J., Finetuned language models are zero-shot learners. arXiv: 2109. 01652 [cs.CL] (2022). sourceCould not verify

  The Introduction (p. 2) discusses large language models’ zero-shot task performance, not the quality of text produced by direct sampling. Searches for grammaticality, simpler sentence constructions, and global incoherence over long spans found no passage in our copy addressing those claims. (LLM, conf=0.75)

  Experimental — accuracy not yet measured
• cites [84] Ouyang L., Training language models to follow instructions with human feedback. arXiv: 2203. 02155 [cs.CL] (2022). sourceCould not verify

  The Introduction (page unavailable) discusses factual errors, biased text, and failure to follow instructions, but not the claimed pattern in directly sampled text. Searches for grammaticality, simpler sentence constructions, and long-span incoherence found no passage in our copy that addresses the claim. (LLM, conf=0.73)

  Experimental — accuracy not yet measured
• cites [85] Rafailov R., “ Direct Preference Optimization: Your Language Model is Secretly a Reward Model” in Advances in Neural Information Processing Systems, vol. 36, Oh A. et al., Eds. (Curran Associates, Inc., 2023), pp. 53728– 53741. sourceCould not verify

  The Introduction discusses why language-model outputs may need steering, but does not describe the claimed grammaticality, preference for simpler sentences, or long-span incoherence. The Additional information section discusses meaningless responses from an unlikelihood-training baseline, not the typical results of directly sampling a large language model. Searches for the claim’s distinctive wording and for long-text incoherence found no supporting passage in our copy; page numbers were not available. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

May combine two claims.

In Evo-generated sequences, TGA and TAA were the most frequent stop codons and TAG the least frequent, matching prior observations in prokaryotes.

“TGA and TAA stop codons appeared most frequently, whereas TAG was least common, consistent with previously observed patterns in prokaryotic genomes (Fig. 6H) (80).”

• cites [80] Korkmaz G, Holm M, Wiens T, Sanyal S, Comprehensive analysis of stop codon usage in bacteria and its correlation with release factor abundance. J. Biol. Chem 289, 30334– 30342 (2014). doi: 10.1074/jbc.M114.606632 25217634 PMC4215218 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Stop codon frequencies across reading frames are a conserved feature of prokaryotic genomes.

“We also examined stop codon frequencies across reading frames, a conserved genomic feature in prokaryotes (79).”

• cites [79] Xu L, Kuo J, Liu J-K, Wong T-Y, Bacterial phylogenetic tree construction based on genomic translation stop signals. Microb. Inform. Exp 2, 6 (2012). doi: 10.1186/2042-5783-2-6 22651236 PMC3466146 sourcePartly supported

  The Discussion reports that related bacterial species often have similar stop-signal ratios in the second and third reading frames, and the Conclusion says signals across all three frames are interrelated. This supports a conserved pattern among related bacteria, but the paper does not establish that stop-codon frequencies are conserved across prokaryotes generally. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

In natural language and image generation, algorithmic methods have been developed that improve output quality beyond sampling from the pretrained model alone.

“Promisingly, in these domains, algorithmic techniques have emerged to improve the quality of generations compared with sampling from the pretrained model alone (83–85).”

• cites [83] Wei J., Finetuned language models are zero-shot learners. arXiv: 2109. 01652 [cs.CL] (2022). sourcePartly supported

  The Introduction (p. 2) describes instruction tuning a pretrained language model, and the Results (p. 4) report that it significantly improves performance over the untuned model on most evaluated NLP tasks, including some generative tasks. The paper does not establish improved quality of open-ended generations compared with sampling from the pretrained model alone, and it does not study image generation. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [84] Ouyang L., Training language models to follow instructions with human feedback. arXiv: 2203. 02155 [cs.CL] (2022). sourcePartly supported

  Section 3.1 describes fine-tuning a pretrained language model using demonstrations, human comparisons, and reward optimization. Section 4.1 reports that labelers preferred the resulting InstructGPT outputs to GPT-3 outputs on the paper’s test prompts. This supports improved natural-language generation over the pretrained model, but the cited paper does not establish the image-generation part of the claim. (LLM, conf=0.87)

  Experimental — accuracy not yet measured
• cites [85] Rafailov R., “ Direct Preference Optimization: Your Language Model is Secretly a Reward Model” in Advances in Neural Information Processing Systems, vol. 36, Oh A. et al., Eds. (Curran Associates, Inc., 2023), pp. 53728– 53741. sourcePartly supported

  The Related Work section says fine-tuning language models with human preferences improves performance on several text-generation tasks, and Section 6.2 reports better summarization results for DPO than for other methods. This supports the claim for natural-language generation, but the paper does not establish it for image generation. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Natural language models frequently have difficulty producing coherent and diverse text over long sequences.

“Natural language models often struggle to maintain coherent and diverse generation over long sequences”

• cites [39] Kaplan J., Scaling Laws for Neural Language Models. arXiv: 2001. 08361 [cs.LG] (2020).Could not verify

  The Introduction (p. 2) says recent models can compose coherent multiparagraph text, while Appendix D.5 (p. 25) examines loss and longer-range correlations across a context. Neither passage addresses whether language models often struggle to maintain coherence and diversity when generating long sequences; searches for those terms found no supporting passage. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [83] Wei J., Finetuned language models are zero-shot learners. arXiv: 2109. 01652 [cs.CL] (2022). sourceCould not verify

  The Introduction (p. 2) discusses difficulty with zero-shot tasks, not maintaining coherence or diversity in long text. The qualitative examples (p. 34) include a story of at least 150 tokens but do not make the claimed general statement. Searches for coherence, diversity, and long-form generation found no passage addressing it. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [84] Ouyang L., Training language models to follow instructions with human feedback. arXiv: 2203. 02155 [cs.CL] (2022). sourceCould not verify

  The Introduction describes failures such as making up facts, producing biased or toxic text, and not following instructions, but does not discuss coherence or diversity over long sequences. Searches for long-form generation, repetition, and output diversity found no passage establishing the claimed difficulty; the References merely list a work on coherent long-form text generation. (LLM, conf=0.72)

  Experimental — accuracy not yet measured
• cites [85] Rafailov R., “ Direct Preference Optimization: Your Language Model is Secretly a Reward Model” in Advances in Neural Information Processing Systems, vol. 36, Oh A. et al., Eds. (Curran Associates, Inc., 2023), pp. 53728– 53741. sourceCould not verify

  The Additional information section says that an unlikelihood-training baseline produced meaningless responses on summarization and dialogue tasks, but it does not discuss whether language models generally struggle with coherence and diversity over long sequences. Searches for coherent and diverse long-sequence generation found no passage addressing that claim; the copy provides no page number for the cited section. (LLM, conf=0.72)

  Experimental — accuracy not yet measured
• cites [87] Kojima T, Gu SS, Reid M, Matsuo Y, Iwasawa Y, “ Large Language Models are Zero-Shot Reasoners” in Advances in Neural Information Processing Systems, vol. 35, Koyejo S., Eds. (Curran Associates, Inc., 2022), pp. 22199– 22213. sourceCould not verify

  The discussion of Few-shot-CoT (p. 2) says language models have struggled with multi-step reasoning, not with maintaining coherence or diversity in long text generation. Searches for long-sequence generation, coherence, diversity, and repetition found no passage in our copy addressing the attributed claim. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Prior work shows that pretraining in multiple sequence-length stages lowers the compute hours needed for long-context models.

“Multistage sequence length pretraining has been shown to reduce the overall number of compute hours required to train long context models (116).”

• cites [116] Xiong W., Effective long-context scaling of foundation models. arXiv: 2309. 16039 [cs.CL] (2023). sourcePartly supported

  Section 4.4 (Training Curriculum; page not available in our copy) compares two-stage sequence-length training with training at the long sequence length throughout and reports about 40% fewer FLOPs with almost no performance loss. That supports reduced computational work, but the paper does not report compute hours, so the claim's specific time-based measure is not established. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Tests of grouped-query attention found little change in final loss, suggesting it could be useful for DNA modeling by lowering inference memory.

“We experimented with grouped-query attention (GQA) (117) and found minimal differences in final loss, suggesting the technique may be suited to DNA sequence modeling, to further reduce memory footprint during inference.”

• cites [117] Ainslie J., GQA: Training generalized multi-query transformer models from multi-head checkpoints. arXiv: 2305. 13245 [cs.CL] (2023). sourceSupported

May combine two claims.

Evo is a genomic foundation model with 7 billion parameters, a context length of 131,072 tokens, and single-nucleotide byte-level tokenization.

“Evo is a genomic foundation model with 7 billion parameters trained with a context length of up to 131,072 tokens, using single-nucleotide, byte-level tokenization.”

May combine two claims.

The compute-optimal token count for Evo 7B was estimated at 250 billion given the FLOP budget; the initial 8192-length pretraining used 300 billion tokens, a 17% offset from the compute-optimal model size.

“We estimate 250 billion to be the compute-optimal number of tokens for Evo 7B given the floating point operation (FLOP) budget, meaning the model was trained at a 17% offset from the compute-optimal model size during the initial 8192 sequence length pretraining phase of 300 billion tokens.”

May combine two claims.

Transformer++ and Mamba showed numerical instability and greater scaling-rate degradation outside the compute-optimal frontier, whereas StripedHyena did not.

“Both Transformer++ and Mamba experienced numerical instability during training and suffered from a higher performance degradation of the scaling rate outside the compute-optimal frontier, in contrast to StripedHyena”

May combine two claims.

Evo is based on the StripedHyena architecture, which combines attention and data-controlled convolutional operators.

“Evo uses a context length of 131,072 tokens and is based on the StripedHyena architecture (24), which hybridizes attention and data-controlled convolutional operators to efficiently process and recall patterns in long sequences.”

• cites [24] Poli M., StripedHyena: Moving Beyond Transformers with Hybrid Signal Processing Models, GitHub (2023); https://github.com/togethercomputer/stripedhyena.Not checked yet

  Not checked in the free run.

May combine two claims.

One of 11 tested Cas9 generations exhibited robust activity and was named EvoCas9-1.

“One of the generations exhibited robust activity, which we named EvoCas9–1.”

May combine two claims.

More conserved sequences tended to be essential, with association strength similar to Evo with gene-only context but weaker than Evo with genomic sequence context.

“We observed that more conserved sequences tended to be essential, with an association strength similar to that of Evo with gene-only context but weaker than that of Evo with genomic sequence context”

May combine two claims.

The predicted EvoCas9-1 backbone resembles SpCas9 but has a more positive surface charge distribution.

“Although the predicted backbone structure of EvoCas9–1 resembles that of SpCas9, the predicted structure of EvoCas9–1 exhibits a more positive surface charge distribution”

May combine two claims.

Concatenating promoter and RBS sequences improved Evo's zero-shot correlation with protein expression to r = 0.61, higher than promoter alone (r = 0.47).

“when concatenating the promoter and RBS sequence together, Evo’s zero-shot likelihoods improved substantially (Spearman r = 0.61); this correlation is also higher than the zero-shot correlation of just the promoter sequence alone (Spearman r = 0.47)”

May combine two claims. The quote only partly states this claim.

TnpA and TnpB proteins detected in Evo-generated sequences varied widely in distance from the nearest training examples, with high ESMFold pLDDT for predicted structures above 40-50% identity and a length distribution closely matching the training set.

“TnpA and TnpB proteins that were detected within these generated sequences varied widely in their distance from the nearest examples in the training set (Fig. 4C), with consistently high ESMFold pLDDT values for predicted structures that were >40 to 50% identity to the training set”

May combine two claims.

The observed transposase activity depended on a putative catalytic tyrosine and on a single-stranded DNA substrate rather than double-stranded DNA, consistent with the known IS200/IS605 TnpA mechanism.

“This activity was also dependent on the presence of a putative catalytic tyrosine and on having a ssDNA substrate instead of double-stranded DNA (dsDNA), consistent with the known mechanism for IS200/IS605 TnpA”

May combine two claims.

Evo's promoter-activity correlations exceed those of GC content (mean r = 0.35) and GenSLM zero-shot likelihoods (mean r = 0.09).

“These correlations also exceed those of the sequence guanine-cytosine (GC) content (mean Spearman r = 0.35) and the zero-shot likelihoods of GenSLM (mean Spearman r = 0.09)”

May combine two claims.

The Evo-designed sgRNA is 91.1% identical to the canonical SpCas9 sgRNA and has secondary-structure differences in the two terminal stem loops, extending stem loops 2 and 3.

“Evo-designed sgRNA is 91.1% identical to the canonical SpCas9 sgRNA and exhibits secondary structure differences in the two terminal stem loops, notably extending the length of stem loops 2 and 3”

May combine two claims.

11 of 24 Evo-generated IS200-like elements and 3 of 24 Evo-generated IS605-like elements showed evidence for both excision and insertion in vitro.

“We observed that 11 out of 24 Evo-generated IS200-like elements and 3 out of 24 Evo-generated IS605-like elements demonstrated evidence for both excision and insertion in vitro”

May combine two claims.

EvoCas9-1 shares 79.9% amino acid identity with the closest Cas9 in the fine-tuning database and 73.1% identity with SpCas9.

“The EvoCas9–1 amino acid sequence shares 79.9% identity with the closest Cas9 in the database of Cas proteins used for model fine-tuning and 73.1% identity with SpCas9.”

May combine two claims.

Providing additional genomic context beyond the gene sequence improves Evo's essentiality prediction, especially from gene-only to 8k context; from 8k to 66k the average performance is comparable, though lower-range examples improve with longer context.

“We also observed that providing the model with additional genomic context beyond the gene sequence results in a substantial improvement in performance, especially from gene-only context to 8k context. From 8k to 66k context, the average predictive performance is comparable, although performance on the lower range of examples does improve with longer context”

May combine two claims.

The 14 active Evo-generated elements use diverse hairpins and encode functional TnpA proteins with sequence identity as low as 67% to the fine-tuning database.

“As a whole, the 14 active elements use a diverse set of hairpins (Fig. 4, E and H, and figs. S25 and S26) and encode functional TnpA proteins with sequence identity as low as 67% to the fine-tuning database.”

May combine two claims.

A CNN on Evo embeddings outperformed ridge regression, and Evo embeddings outperformed one-hot embeddings across both architectures.

“The CNN architecture substantially outperformed ridge regression across both embeddings, and the Evo embeddings substantially outperformed one-hot embeddings across both architectures”

May combine two claims.

Functional IS605-like Evo-generated elements contain putative TnpB coding sequences and sequences with significant matches to a covariance model built from known ωRNAs.

“The functional IS605-like elements, which contain putative TnpB coding sequences, also contain sequences with significant matches (cmsearch E value < 0.001) to a covariance model constructed from known ωRNAs”

May combine two claims.

Compared with HyenaDNA, Evo uses an improved hybrid design and is scaled to 1000× larger model size and 100× more data.

“Compared with HyenaDNA (33), a previous generation of DNA models leveraging a Hyena architecture (34), Evo is based on an improved hybrid design and scaled to 1000× larger model size and 100× more data.”

• cites [33] Nguyen E., “ HyenaDNA: Long-Range Genomic Sequence Modeling at Single Nucleotide Resolution” in Advances in Neural Information Processing Systems, vol. 36, Oh A. et al., Eds. (Curran Associates, Inc., 2023), pp. 43177– 43201. sourceNot checked — no full text of the source
• cites [34] Poli M., Hyena Hierarchy: Towards Larger Convolutional Language Models. arXiv: 2302. 10866 [cs.LG] (2023). sourceNot checked — no full text of the source

May combine two claims.

Evo is a genomic foundation model trained on hundreds of billions of DNA tokens spanning prokaryotic evolutionary diversity, and it can perform prediction and generation from molecules up to whole genomes.

“Evo is a genomic foundation model trained on hundreds of billions of DNA tokens across the evolutionary diversity of prokaryotic life, capable of prediction and generation tasks at the scale of individual molecules, molecular complexes, systems, and even whole genomes.”

May combine two claims.

Natural and Evo-generated sequences show similar patterns of coding organization, with nearby coding sequences typically on the same strand.

“When visualized, both natural and generated sequences display similar patterns of coding organization (Fig. 6C), with sequences in close proximity typically found with the same strand orientation”

May combine two claims.

Evo is built on a state-of-the-art hybrid architecture and supports single-nucleotide-resolution language modeling with a context length of 131,072.

“Based on a state-of-the art hybrid model architecture, Evo enables single-nucleotide-resolution language modeling at a context length of 131,072.”

May combine two claims.

Evo was used to sample 16 sequences of about 1 Mb each, exceeding its 131 kb context length by more than sevenfold.

“We used Evo to sample 16 sequences each containing ~1 Mb, representing more than seven times the model’s context length of 131 kb.”

May combine two claims.

Evo made accurate zero-shot predictions on varied protein, ncRNA, and regulatory DNA fitness or expression tasks, equaling or exceeding specialized models, and captured how single-gene mutations can influence organism-level fitness.

“Evo accurately performed zero-shot prediction across diverse fitness or expression prediction tasks on proteins, ncRNAs, or regulatory DNA that matches or outperforms specialized models while also understanding how mutations to individual genes can affect broader organismal fitness.”

May combine two claims.

Evo was applied to generate CRISPR-Cas proteins with their guide RNAs, multicomponent transposable systems, and megabase-scale sequences resembling real genome architecture.

“As a multimodal generative model, we use Evo to generate CRISPR-Cas proteins and their noncoding guide RNAs, multicomponent transposable systems, and megabase-long sequences that recapitulate the architecture of real genomes.”

May combine two claims.

Species-specific Evo generations correlated strongly with their natural reference sequences, and TUDs were accurate enough to reconstruct phylogenetic relationships among the generated sequences.

“We found strong correlations between species-specific generations and their corresponding natural reference sequences, with TUDs sufficiently accurate to reconstruct natural phylogenetic relationships among the generated sequences (Fig. 6, F and G).”

• AssumptionRests on an assumption: TUDs computed on Evo-generated sequences can reconstruct phylogenetic relationships among those sequences.

May combine two claims.

The paper reports the first scaling-law study of DNA pretraining across multiple architectures, finding that StripedHyena surpassed several baselines, including Transformers.

“We conducted the first scaling laws analysis of DNA pretraining across several architectures, where we observed StripedHyena outperforming several baseline architectures, including Transformers.”

May combine two claims.

Evo-generated sequences have coding densities nearly equal to natural genomes and substantially higher than random sequences.

“Notably, Evo generated sequences have nearly the same coding densities as natural genomes, and substantially higher than that of random sequences (Fig. 6B).”

• AssumptionRests on an assumption: CheckM metrics computed on Evo-generated sequences are comparable to those computed on natural bacterial genomes.

The quote only partly states this claim.

Evo can show similar long-sequence generation limitations.

“Evo can demonstrate similar properties.”

May combine two claims.

Evo's pretraining used 300 billion prokaryotic tokens, a tiny fraction of the petabytes of public genomic data.

“We pretrained Evo on a dataset of 300 billion prokaryotic tokens, which represents a miniscule portion of petabytes of publicly available genomic data.”

May combine two claims.

At genome scale, Evo produces megabase-long sequences with high-level genome organization, yet it often fails to include key marker genes like complete rRNA sets.

“At the genome-scale, Evo generates megabase-long sequences that demonstrate a high-level understanding of genome organization, but it struggles to include key marker genes, such as full sets of rRNAs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In some cases, adding an extra random nucleotide to the Cas9 prompt improved the quality of generated sequences.

“we found that, in some instances, adding an additional random nucleotide character would improve the quality of generations.”

May combine two claims.

Evo can generate synthetic CRISPR-Cas molecular complexes and transposable systems, and experimental validation confirmed functional activity for Evo-generated CRISPR-Cas complexes and IS200 and IS605 transposable systems.

“We experimentally validated the functional activity of Evo-generated CRISPR-Cas molecular complexes as well as IS200 and IS605 transposable systems”

May combine two claims.

EvoCas9-1 was generated from only 11 designs, which represents a robust success rate given the complexity of Cas9's multistep mechanism.

“EvoCas9–1 was generated from just 11 code-signs, representing a robust success rate given the complexity of Cas9’s multistep mechanism”

Probably not a substantial claim: may not pass S3 (load-bearing).

Model hybridization improves scaling performance on language modeling for both standalone Hyena and Transformer architectures.

“Model hybridization, first proposed to address shortcomings of state-space models (30–32), has recently been shown to improve scaling performance on language modeling of both standalone Hyena and Transformer architectures (24).”

• cites [24] Poli M., StripedHyena: Moving Beyond Transformers with Hybrid Signal Processing Models, GitHub (2023); https://github.com/togethercomputer/stripedhyena.Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Evo is a foundation model designed to capture the multimodality of the central dogma and the multiscale nature of evolution.

“Evo is a foundation model that is designed to capture two fundamental aspects of biology: the multimodality of the central dogma and the multiscale nature of evolution.”

Probably not a substantial claim: may not pass S4 (based).

Evo learns covariation patterns involving any genetic elements within its context window.

“Because Evo learns covariation patterns involving any genetic elements within its context window, the model should understand interactions between encoded protein and ncRNA molecules.”

May combine two claims.

Evo demonstrates a robust and general understanding of constitutive protein coding sequences, ncRNA sequences, and regulatory elements.

“Evo demonstrates a robust and general understanding of the constitutive protein coding sequences, ncRNA sequences, and regulatory elements.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The smallest minimal bacterial genomes are approximately 580 kb long.

“For comparison, the smallest “minimal” bacterial genomes are ~580 kb in length (76).”

• cites [76] Blanchard A, Bébéar C, The evolution of Mycoplasma genitalium. Ann. N. Y. Acad. Sci 1230, E61– E64 (2011). doi: 10.1111/j.1749-6632.2011.06418.x 22417108 sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

The authors used a negative control of 1000 uniformly random DNA sequences of length 131,072, computing CheckM coding densities on them.

“As a negative control, we generated 1000 sequences of length 131,072 in which the four DNA base pairs were sampled uniformly at random.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

To account for different start codons and produce a more conservative identity estimate, TnpA and TnpB alignments were trimmed to the middle 80% of each sequence, end gaps were removed, and amino acid percent identity was recalculated as trimmed percent identity.

“To account for different start codons and to generate a more conservative percentage identity estimate, these alignments were trimmed to the middle 80% of each sequence, end gaps were trimmed, and the amino acid percent identity was recalculated, which we called a “trimmed percent identity.””

Probably not a substantial claim: may not pass S3 (load-bearing).

CheckM computes coding density as one metric of genome quality.

“CheckM uses Prodigal (140) to identify coding sequences and computes the coding density as one metric of genome quality.”

• cites CheckM (77)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

To select Cas9 candidates for experimental validation, generated Cas9 ORFs with more than 90% sequence identity to a training sequence were filtered out.

“To select candidates for experimental validation, Cas9 generations with an ORF sequence identity higher than 90% to a training sequence were first filtered out.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The authors used the linear position of a gene in the reference genome as a control value for predicting gene essentiality.

“As an additional control, we also used the gene’s linear position in the reference genome as the value with which to predict essentiality.”

Classified as probably not a substantial claim.

CheckM computes basic genome quality metrics by comparing a long DNA sequence's properties with those of known bacterial genomes.

“We evaluated these generations with CheckM (77), a tool that computes basic genome quality metrics based on whether a given long DNA sequence has similar properties as known bacterial genomes.”

• cites CheckM (77)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

CheckM searches for genes that are highly conserved across much of prokaryotic diversity.

“CheckM will also search for the presence of genes that are highly conserved across much of prokaryotic diversity.”

• cites CheckM (77)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

CheckM uses Prodigal to identify coding sequences.

“CheckM uses Prodigal (140) to identify coding sequences and computes the coding density as one metric of genome quality.”

• cites CheckM (77)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Prodigal (140)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

For final synthesis and experimental validation, a different 60 bp filler sequence was used for ISHp608-like candidates to eliminate a primer-binding site containing TTAC, the canonical ISHp608 target site.

“For final synthesis and experimental validation, a different 60 bp filler sequence was used for the ISHp608-like candidates compared to the ISStin10-like, ISSpn6-like, or ISDge10-like sequences to eliminate a primer-binding site containing a TTAC, which is the canonical ISHp608 target site.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The authors assessed the statistical significance of AUROC using a permutation-based null distribution with 100,000 permutations.

“We assessed statistical significance of the AUROC with a permutation-based method in which a null distribution is constructed by permuting the binary labels and recomputing the subsequent AUROC.”

• Also stated in: § Gene essentiality prediction · plain

  “We performed 100,000 permutations to construct this null distribution.”

Classified as probably not a substantial claim.

To make Spearman correlations comparable across the promoter-alone, RBS-alone, and promoter-RBS settings, the correlation was computed over all 12,243 examples, with sequences duplicated in the single-component settings.

“To ensure that the Spearman correlation is comparable across these settings, we computed the correlation over all 12,243 examples (which involves duplicating sequences in the promoter-alone or RBS-alone settings).”

Probably not a substantial claim: may not pass S3 (load-bearing).

The authors used gene conservation as an additional control.

“We also used the conservation of a gene as another control.”

• AssumptionThe authors state this as an assumption: A greater number of significant mmseqs hits corresponds to higher conservation.
• AssumptionThe authors state this as an assumption: Higher conservation corresponds to greater essentiality.

Probably not a substantial claim: may not pass S3 (load-bearing).

The authors used a positive control of 100 randomly selected GTDB bacterial genomes, computing CheckM coding densities for 131,072 bp crops from them.

“As a positive control, we randomly selected 100 bacterial genomes from GTDB and used CheckM to compute the coding densities for 131,072 bp crops from these genomes.”

• cites GTDBNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites CheckM (77)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

For IS200/IS605 fine-tuning, the data loader was modified so each training sample began with the first token of the IS200/IS605 sequence, and shorter sequences were padded to the remaining context without padding contributing to the loss.

“We also modified the data loader such that each sample provided to the model during training would begin with the first token of the IS200/IS605 sequence and, if a sequence was shorter than the context length, we padded the sequence to the remaining context (where padding did not contribute to the loss computation), similar to the strategy described for CRISPR-Cas9 systems above.”

Probably not a substantial claim: may not pass S3 (load-bearing).

If a model relied only on positional information to predict essentiality, its performance would resemble that of the linear-position control.

“If a model were simply using positional information to make essentiality predictions, the performance would be similar to this control.”