Genome Modelling and Design across All Domains of Life with Evo 2

• p. 13 A.3.4 Stop codon analysis across genetic codes

  Thalassiosira pseudonana (GCA_000149405.2, ciliate code), and Tetrahymena thermophila (GCA_000189635.1, ciliate code)

• p. 4 Fig. 2e, species legend

  P. tetraurelia

• p. 17 Extended Data Fig. 3f, panel titles

  P. tetraurelia

• p. 4 Fig. 2e legend/labels

  P. tetraurelia, Paramecium tetraurelia

Judge's note: The Fig. 2e image (main p.4) has a legend listing A. thaliana, H. sapiens, M. pneumoniae, P. tetraurelia and T. thermophila. Extended Data Fig. 3f (main p.17) plots P. tetraurelia and T. thermophila for both the WT and the recoded genomes. Supplementary A.3.4 lists T. pseudonana in both the 5-species list and the recoding paragraph. A.3.3 gives P. tetraurelia a separate accession (GCA_000165425.1), so the methods name the wrong species for what is plotted.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 8 Results, 'Genome-scale generation'

  we generated ten 580-kb sequences that we annotated with Prodigal

• p. 24 A.5.3 Mycoplasma genitalium genome generation

  Using Evo 2 40B, we generated 35 sequences of 580 kb each with temperature 1.0 and top-k of 4.

• p. 8 Genome-scale generation

  we gener-ated ten 580-kb sequences

• p. 9 Fig. 5 legend g

  Example Evo 2-generated approximately 600-kb DNA sequence

Only the authors can say what is right here.

Judge's note: Both passages describe the same experiment: a 10.5-kb M. genitalium prompt, 580-kb outputs, Prodigal annotation and hmmscan/Pfam. Neither mentions taking a subset. The '~600-kb' in the Fig. 5g legend is qualified as approximate and is not a defect in itself. The count (ten vs 35) is the real discrepancy and needs the authors to resolve it.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 20 Extended Data Fig. 4 caption, panel (b)

  “near” is defined as within 5 bp or 3 bp to donor or acceptor sites, respectively

• p. 18 A.3.13 ClinVar variant effect prediction

  located within 3 base pairs of the nearest splice donor site or within 5 base pairs of the nearest splice acceptor site, reflecting the asymmetric importance of donor versus acceptor consensus sequences

• p. 20 Extended Data Fig. 4b legend

  within 5 bp or 3 bp to donor or acceptor sites, respectively

• p. 18 A.3.13

  within 3 base pairs of the nearest splice donor site or within 5 base pairs of the nearest splice acceptor

Only the authors can say what is right here.

Judge's note: Page 20 image: the Extended Data Fig. 4b caption defines near as 5 bp to donor / 3 bp to acceptor. supp-02 page 18 (A.3.13 ClinVar variant effect prediction) defines the same ClinVar near/far stratification shown in Extended Data Fig. 4b as 3 bp donor / 5 bp acceptor, adding 'reflecting the asymmetric importance of donor versus acceptor consensus sequences'. I searched the article and supplement and found no second, differently defined stratification, so the two statements describe one analysis with the assignments reversed; which is correct can only be settled by the authors.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 29 A.6.8 Sequence generation and decoding strategy

  total designed length L of either 1024 or 4996 bp

• p. 29 A.6.9 Cell-type specific scoring function

  32 or 128 bins for a 4096 bp generated region

• p. 29 Section A.6.9, Cell-type specific scoring function

  resulting in 8 or 32 bins for a 1024 bp generated region, and 32 or 128 bins for a 4096 bp generated region

• p. 29 A.6.8

  either 1024 or 4996 bp

Judge's note: Read the full page image supp-02 p029: A.6.8 prints "chunk size of C = 128 bp and total designed length L of either 1024 or 4996 bp"; A.6.9 prints "output resolution was 128 bp for Enformer and 32 bp for Borzoi, resulting in 8 or 32 bins for a 1024 bp generated region, and 32 or 128 bins for a 4096 bp generated region". The quoted bin counts follow from 4096 (4096/128 = 32, 4096/32 = 128); 4996 does not give them and is not an integer multiple of the 128 bp chunk, while supp-02 A.6.1 (layout text line 1312) states "we assume that C is constant and that L is an integer multiple of C". The same page also refers to "the 1 kb and 4 kb sequence designs". No passage anywhere describes a separate 4996 bp design, so the two values are two readings of one quantity.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 32 Reporting Summary, Life sciences study design, 'Replication'

  We created duplicate clonal cell lines for all HEK293T and K562 designs (Figure sS10H-L).

• p. 2 Reporting Summary, 'Replication'

  (Figure sS10H-L)

• p. 2 Life sciences study design, Replication

  We created duplicate clonal cell lines for all HEK293T and K562 designs (Figure sS10H-L).

• p. 11 Section 'Designing mammalian chromatin patterns'

  (Fig. 6i,j and Extended Data Fig. 11b–h)

Judge's note: Searching the article and supplement text found no figure labelled S10 (no 'Figure S' or 'Supplementary Fig' anywhere). The Extended Data Fig. 11 caption (main p.30) describes panels d–h as the designs integrated into both HEK293T and K562. The pointer also contains a stray 's' ('sS10'), visible in the page image.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 3 Results, 'Evo 2 architecture, training, and data'

  initial pretraining at shorter context lengths improves both efficiency and overall model quality9–11

• p. 12 Reference list, entry 11

  Liu, S. J. et al. In vivo perturb-seq of cancer and microenvironment cells dissects oncologic drivers and radiotherapy responses in glioblastoma. Genome Biol. 25, 256 (2024).

• p. 3 Section 'Evo 2 architecture, training, and data', second paragraph

  This matches best practice in natural language, in which initial pretraining at shorter context lengths improves both efficiency and overall model quality9–11.

• p. 12 Reference list, ref. 11

  In vivo perturb-seq of cancer and microenvironment cells dissects oncologic drivers and radiotherapy responses in glioblastoma

Only the authors can say what is right here.

Judge's note: I checked the citation on p3 and the reference list on p12. Refs 9 and 10 are language-model papers on long-context training and Llama 3. Ref. 11 is a glioblastoma Perturb-seq study and does not bear on language-model pretraining. Ref. 11 is cited nowhere else in the text. The citation looks wrong; only the authors can supply the intended reference.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 19 Extended Data Fig. 4b, Noncoding SNVs and Noncoding non-SNVs panels

  N = 2,743

• p. 5 Fig. 3 legend

  noncoding (c; n = 34,761 SNVs, n = 3,894 non-SNVs)

Only the authors can say what is right here.

Judge's note: Read the page 19 image (no usable text layer, so judged from the image): the far/near panels for Noncoding SNVs and for Noncoding non-SNVs show the same pair of N values. The sum 2,743 + 1,074 = 3,817 is close to the 3,894 noncoding non-SNVs given in the Fig. 3 legend but an order of magnitude below the 34,761 noncoding SNVs. The coding row, by contrast, carries its own values (13,287 far / 1,026 near, with the near coding non-SNV panel omitted as the caption explains for the 22 such variants). Only the authors can supply the correct N for the noncoding SNV panels.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 21 Extended Data Fig. 5 caption, panel (d)

  (d) Comparison of AUROCs for the supervised classification of BRCA1 SNVs using embeddings from Evo 2 40B block 20, for pooling functions that use different window sizes around the variant site to average embeddings.

• p. 21 Extended Data Fig. 5, panel d x-axis

  Evo 2 40B block index

• p. 21 Extended Data Fig. 5, panel e x-axis

  Pooling function window size (nt)

• p. 21 Extended Data Fig. 5 caption, (d) and (e)

  (d) Comparison of AUROCs for the supervised classification of BRCA1 SNVs using embeddings from Evo 2 40B block 20, for pooling functions

Judge's note: On the page 21 image, panel d's x-axes read 'Evo 2 40B block index' / 'Evo 1 block index' and panel e's x-axis reads 'Pooling function window size (nt)'. The caption gives the opposite descriptions. The main text (p.6) cites 'Extended Data Fig. 5d' for layer selection, which matches the plotted panels, not the caption.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 7 Fig. 4f, right-hand motif logo panel

  f/24109

• p. 7 Fig. 4f, left-hand promoter activation tracks

  f/24019

• p. 7 Fig. 4f labels

  FOXE1 motif f/24109

Judge's note: I checked the Fig. 4f crop (p7). The legends for all three promoter tracks list f/24019 for the FOXE1/AAACAA peaks, but the FOXE1 motif logo is captioned f/24109. This looks like a digit transposition in a feature ID.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 23 A.4.8 Exon-intron features

  The activations for the same set of features were collected for the PDK3 gene loci of the woolly mammoth genome

• p. 7 Fig. 4g, gene track label

  PDK4

• p. 23 A.4.8

  PDK3 gene loci of the woolly mammoth

Only the authors can say what is right here.

Judge's note: Both read from the images (supp-02 p.23 crop; main p.7 Fig. 4g). Only one mammoth locus is described and shown, and these are the only two mentions of PDK in the files. Which gene name is correct cannot be settled from the files.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 8 Results, 'Genome-scale generation'

  with varying degrees of sequence similarity to natural genes (Fig. 5d and Supplementary Table 7)

• p. 43 Appendix C.3, Supplementary Table 7 caption

  Supplementary Table 7 | Comparing Evo 2 7B base and data ablation model on zero-shot SpliceVarDB classification

• p. 44 Supplementary Table 10 caption

  Supplementary Table 10 | BLASTp results for genes from Evo 2-generated mitochondria

• p. 8 Section 'Genome-scale generation', mitochondrial generation paragraph

  with varying degrees of sequence similarity to natural genes (Fig. 5d and Supplementary Table 7) while maintaining proper synteny (Fig. 5e)

Judge's note: I checked the main text (p8) and the captions of supp-02 Table 7 (p43) and Table 10 (p44). Table 10 is 'BLASTp results for genes from Evo 2-generated mitochondria' and gives sequence ID %. Table 7 is unrelated, so the pointer is wrong.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 9 Fig. 5 legend, panel l

  are visualized alongside predicted homologous yeast gene, exon, promoter and tRNA annotations

• p. 9 Fig. 5l, track labels

  Intron

• p. 26 Extended Data Fig. 9 caption, panel (i)

  Distribution of genes, introns, tRNAs, and promoters on an Evo 2-generated S. cerevisiae sequence

Judge's note: In the page 9 image, both the native S. cerevisiae chromosome III panel and the Evo 2-generated panel carry exactly four annotation rows: Gene, Intron, Promoter, tRNA. No exon track is drawn. The Extended Data Fig. 9 caption on page 26 likewise lists 'genes, introns, tRNAs, and promoters'. The legend names a track type that is not plotted.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 11 Results, 'Designing mammalian chromatin patterns'

  1 out of 24 designs (4%) had greater than threefold differential accessibility (Fig. 6i and Extended Data Fig. 11b)

• p. 10 Fig. 6i panel title

  Same pattern, different cell types

• p. 10 Fig. 6j panel title

  Different patterns, different cell types

• p. 11 Designing mammalian chromatin patterns

  had greater than threefold differential accessibility (Fig. 6i and Extended Data Fig. 11b)

Judge's note: Page 11 crop gives the sentence; page 10 image shows panel i titled 'Same pattern, different cell types' (AUROC 0.98 HEK293T / 0.97 K562, same peaks in both) and panel j titled 'Different patterns, different cell types' (flat HEK293T, single K562 peak, AUROC 1.00). The Fig. 6 legend states that integration into both cell lines 'enables the design of identical patterns across both cell types (i) or of cell-type-specific accessibility profiles (j)'. The earlier sentence in the same paragraph cites 'Fig. 6i,j' jointly, so the authors do distinguish the panels.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 12 Reference 36

  Mol. Cells 38

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: that this 2010 paper appeared in Molecular Cell (Mol. Cell) rather than the separate journal Molecules and Cells (Mol. Cells). The page 12 image confirms the entry is printed as 'Mol. Cells 38, 576-589 (2010)' as quoted; the volume and pages given are consistent with a Mol. Cell citation.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 31 B.1

  tRNA adaptation index (tAI) mutations A.3.2

• p. 31 B.1

  tRNA adaptation index (tAI) mutations A.3.2.

• p. 13 A.3.5 heading

  A.3.5. Translational codon ramp analysis

• p. 31 B.1 Codon usage analysis

  higher versus lower tRNA adaptation index (tAI) mutations A.3.2

Judge's note: Read the full page image supp-02 p031: B.1 (Codon usage analysis) describes the codon-ramp/tAI mutation comparison and cites section A.3.2 as a hyperlinked cross-reference. Read crop C050-c and the surrounding page supp-02 p013: section A.3.5 is headed "Translational codon ramp analysis" and describes exactly this method (tAI per codon, synonymous mutations with higher vs lower tAI, average log-likelihood change). The contents list (layout text line 86) shows A.3.2 is "Effect of prokaryotic mutations on Evo 2 likelihoods", a different analysis. Pointer is wrong; the data are unaffected.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 18 Extended Data Fig. 3 legend, (e)

  while red lines indicate synonymous codon mutations with a lower tAI than the reference sequence

• p. 18 Extended Data Fig. 3e panel legend

  Lower tAI

Judge's note: Page 18 crop gives the caption wording; page 17 image shows panel e with a two-entry key reading 'Lower tAI' and 'Higher tAI', both drawn in shades of blue, with no red line in the panel. The colour named in the caption does not match the plotted key.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 24 A.5.4 Yeast chromosome generation

  Gene prediction was performed using GeneMark-S with Prot-hint based on the Fungi proteins from OrthoDB v12.0.

• p. 26 Extended Data Fig. 9 caption, panel (j)

  Distribution of gene lengths and pLDDTs of Evo 2-generated and natural S. cerevisiae genes, annotated by GeneMark-ES.

Only the authors can say what is right here.

Judge's note: Read the crop and full page supp-02 p024: A.5.4 plainly reads "GeneMark-S with Prot-hint" for the Evo 2-generated S. cerevisiae chromosomes. Read the crop and full page main p026: Extended Data Fig. 9j reads "annotated by GeneMark-ES" for the same generated S. cerevisiae genes, and main p009 (Fig. 5l) labels the tracks only "Gene called by GeneMark". Both passages describe annotation of the same generated yeast sequences, so the tool name is inconsistent; only the authors can say which variant was run (the ProtHint-based workflow is normally GeneMark-ES/EP, but that is outside knowledge and the internal inconsistency alone is enough). Rests on outside knowledge only for the claim about which variant pairs with ProtHint; the S vs ES mismatch is visible in the files.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 29 A.6.9 Cell-type specific scoring function, item 1

  1. Accessible in HEK293T and inaccessible in K562 (HEK+/K562−):

• p. 29 A.6.9 Cell-type specific scoring function, item 2

  2. Accessible in HEK293T and inaccessible in K562 (HEK+/K562−):

• p. 28 Section A.6.7, list of designed patterns

  (2) inaccessible in HEK293T and accessible in K562 (HEK−/K562+)

Judge's note: Read the full page image supp-02 p029: item 1 is f_total = f_Enformer,HEK293T + f_Borzoi,HEK293T + 0.5 f_Enformer,K562 + 0.5 f_Borzoi,K562; item 2 is f_total = 0.5 f_Enformer,HEK293T + 0.5 f_Borzoi,HEK293T + f_Enformer,K562 + f_Borzoi,K562; both headings are word-for-word identical. Read supp-02 p028 (A.6.7), which defines four patterns: "(1) accessible in HEK293T and inaccessible in K562 (HEK+/K562−), (2) inaccessible in HEK293T and accessible in K562 (HEK−/K562+), (3) uniformly inaccessible (HEK−/K562−), and (4) uniformly accessible (HEK+/K562+)", and item 3 of A.6.9 covers both uniform cases. Nothing elsewhere in the supplement states that two weightings were used for one pattern, so the duplicated heading is a label error; the mirrored weights indicate item 2 is the HEK−/K562+ case.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 32 Reporting Summary, Life sciences study design, 'Replication'

  we created triplicate clonal cell lines for "square wave long," "square wave medium," and "Morse code LO". We created duplicate clonal cell lines for "square wave medium" and "Morse code EVO2".

• p. 2 Reporting Summary, 'Replication'

  triplicate clonal cell lines for "square wave long," "square wave medium," and "Morse code LO". We created duplicate clonal cell lines for "square wave medium" and "Morse code EVO2"

• p. 2 Life sciences study design, Replication

  we created triplicate clonal cell lines for "square wave long," "square wave medium," and "Morse code LO". We created duplicate clonal cell lines for "square wave medium" and "Morse code EVO2".

• p. 26 Section A.6.3, list of designed peak patterns

  Short rectangular wave: Alternating between 384 bp of open chromatin and 1152 bp of closed chro-matin.

Only the authors can say what is right here.

Judge's note: Checked in the page images (main p.32 and supp-01 p.2): the text is identical in both. Supplementary A.6.3 lists six designs: long square, medium square, short rectangular, LO, ARC and EVO2. As printed, the replicate statements cover medium twice and short not at all. No other replicate information resolves this, so only the authors can say which design was made in duplicate.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 8 A.1.6 Midtraining phase: Needle-in-a-haystack evaluation

  at depths corresponding to 10% through 90%, at intervals of 10%, of the total haystack length

• p. 2 Fig. 1i, 'Depth into context (%)' axis

  100

• p. 15 Extended Data Fig. 1d, 'Depth into context (%)' axis

  100

• p. 8 Section A.1.6, null-distribution paragraph

  at depths into context of 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, and 100% of the haystack sequence length

Judge's note: Fig. 1i (main p.2) shows 10 depth rows (10–100), each with values. The null paragraph on supp-02 p.8 lists ten depths including 100% and computes 8×10×13 = 1,040, which is consistent with ten depths. The '10% through 90%' sentence therefore misdescribes the depths evaluated. Extended Data Fig. 1d was not opened; Fig. 1i alone is enough.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 14 A.3.6 Zero-shot protein fitness prediction

  we also compiled a set of 18 DMS datasets corresponding to proteins from viruses that infect humans

• p. 14 A.3.6, end of the same list

  a dengue virus NS5 DMS by Suphatrakul et al., and an influenza PB1 DMS by Li et al.

Only the authors can say what is right here.

Judge's note: Counted from the supp-02 p.14 image: HCV pol; HA (Wu); NP and PB1 (Doud); PA; NA; TAT; REV; env (Haddox); env (Duenas-Decamp); HA (Lee); env (Haddox); PB2; Zika env; SARS-CoV-2 RBD; coxsackievirus capsid; AAV2 capsid; Mpro; dengue NS5; PB1 (Li). That is 20 items. The other counts in the same section (nine prokaryotic, six human) match their lists. The Extended Data Fig. 2b caption gives no count. Only the authors can say whether 18 or the list is correct.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send