Verified Rigor RecordRigor Reviewed

Rigor-Reviewed Manuscript

Genome modelling and design across all domains of life with Evo 2.

Name: Verified Rigor Record — Genome modelling and design across all domains of life with Evo 2.
Item: Genome modelling and design across all domains of life with Evo 2.
Rating: 89
Author: Αlpha¹

Brixi G, Durrant MG, Ku J, Naghipourfar M, Poli M, Sun G, Brockman G, Chang D, Fanton A, Gonzalez GA, King SH, Li DB, Merchant AT, Nguyen E, Ricci-Tam C, Romero DW, Schmok JC, Taghibakhshi A, Vorontsov A, Yang B, Deng M, Gorton L, Nguyen N, Wang NK, Pearce MT, Simon E, Adams E, Amador ZJ, Ashley EA, Baccus SA, Dai H, Dillmann S, Ermon S, Guo D, Herschl MH, Ilango R, Janik K, Lu AX, Mehta R, Mofrad MRK, Ng MY, Pannu J, Ré C, St John J, Sullivan J, Tey J, Viggiano B, Zhu K, Zynda G, Balsam D, Collison P, Costa AB, Hernandez-Boussard T, Ho E, Liu MY, McGrath T, Powell K, Pinglay S, Burke DP, Goodarzi H, Hsu PD, Hie BL.

Paper source

Open DOI

Αlpha¹ Editorial

This rigor review was examined and confirmed by Αlpha¹ Editorial · July 5, 2026

Αlpha¹ Rigor Review

Rigor Reviewed

PASS

Rigor holds up · Computational Study

Eight dimensions — click a segment for details

Critical

High-priority

Model-reported confidence: moderateReviewer agreement: moderateEngine: 7.5.0Source: full text

This Αlpha¹ Rigor Review uses AI Rigor Reviewers trained on a curated corpus of high-fidelity and retracted papers, with expert supervision and curation. It can still make mistakes; verify each finding against the source before relying on it.

The paper presents Evo 2, a large genomic language model, with strong open science practices (data/code fully shared) and generally transparent reporting. However, it has significant reporting gaps regarding cell line authentication, mycoplasma testing, and ethics approvals for the wet-lab validation experiments. Additionally, a substantial number of references (13) could not be found in registries, warranting verification.

Evaluated as a computational study with minor wet-lab validation. Two of eight dimensions (biological variables and ethical approvals) had reviewer disagreement on applicability; I weighted the evidence and applied best-practice expectations. Statistics coverage was limited: 0 of 5 reported tests were machine-verifiable (threshold p-values, resampling methods); no errors were found but numerical precision is not independently confirmed.

12 major claims checked against the paper's own evidence: all adequately supported.

Found 5 reported tests, but none could be recomputed (missing degrees of freedom or sample size).

3 integrity concerns flagged (0 high).

Rigor

Eight Dimensions5 pass · 1 fail

12 major claims3 partial9 supported

The introduction and background sections lay out the motivation: tools for sequencing and editing have advanced, but understanding genomic complexity remains difficult. The paper cites the success of previous models (Evo 1, ProGen) and identifies limitations (e.g., prokaryotic-only training, short context windows) that Evo 2 aims to address. The premise is logical and well-supported.

Sub-criteria

prior work citedADEQUATE

premise rationaleADEQUATE

limitations addressedADEQUATE

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 90% conf

The paper establishes a clear scientific foundation by citing prior work (Evo 1, other genomic language models), providing a logical rationale for the study, and addressing limitations of previous approaches.

Evidence

direct quote[Introduction, paragraph 1]

“We previously demonstrated that machine learning models trained on prokaryotic genomic sequences can model the function of DNA, RNA and proteins...”

direct quote[Introduction, paragraph 2]

“We reasoned that training a model with these capabilities would require data spanning the full spectrum of biological diversity to discover emergent properties similar to those found in other fields.”

direct quote[Results, 'Human variant effect prediction']

“Genomic language models have previously struggled in eukaryotic variant effect prediction, lagging considerably behind species-specific models that use multiple sequence alignments”

PASSReviewer 2· 95% conf

The premise is well-established: genomic sequences encode biological function, and large-scale pre-training can capture this information. The paper extensively cites prior work on genomic language models (Evo 1, Nucleotide Transformer, etc.) and provides a clear rationale for a multi-species, multi-domain model.

Evidence

direct quote[Introduction, paragraph 1-2]

“AI models that learn information from genomic sequences across diverse organisms have increasingly advanced prediction and design capabilities”

direct quote[Introduction, paragraph 2]

“We reasoned that training a model with these capabilities would require data spanning the full spectrum of biological diversity”

direct quote[Introduction, paragraph 3]

“We previously demonstrated that machine learning models trained on prokaryotic genomic sequences can model the function of DNA, RNA and proteins... Here we present Evo 2... extend the sequence modelling paradigm to the scale and complexity of eukaryotic genomes”

PASSReviewer 3· 95% conf

The paper thoroughly cites prior work on biological foundation models (e.g., Evo 1, Nucleotide Transformer) and clearly establishes the rationale for scaling to genome-wide and multi-domain data.

Evidence

direct quote[Introduction, paragraph 2]

“We previously demonstrated that machine learning models trained on prokaryotic genomic sequences can model the function of DNA, RNA and proteins... Here we present Evo 2, a biological foundation model trained on a representative snapshot of genomes spanning all domains of life.”

direct quote[Introduction, paragraph 1]

direct quote[Introduction, paragraph 2]

“We extend the sequence modelling paradigm to the scale and complexity of eukaryotic genomes through advances in data curation, model architecture, large-scale pre-training, advanced interpretability methods and inference-time prediction and generation approaches.”

Randomization, blinding, and power analysis are not applicable for this computational study. Inclusion/exclusion criteria for training data are clearly stated (e.g., excluding eukaryotic viruses). Outlier handling is embedded in data curation and modeling choices. Controls are present as baseline comparisons (e.g., Evo 1, transformers). Independent replication is not applicable as this is a single model, but multiple benchmark evaluations serve as validation. Replicate distinction is not applicable as there are no biological replicates. The design is appropriate for the study type.

Sub-criteria

controlsADEQUATE

power analysisN/A

blinding levelsN/A

outlier handlingINADEQUATE

randomization unitN/A

inclusion exclusionADEQUATE

randomization methodN/A

replicate distinctionN/A

independent replicationN/A

Per reviewer (3) — the votes behind the verdict

WARNReviewer 1· 75% conf

The study design is reasonable for a computational paper, but several elements are missing or insufficiently reported: outlier handling, independent replication, and a formal power analysis for dataset selection are not addressed.

Evidence

direct quote[Results, 'Evo 2 architecture, training, and data']

“This new training dataset, which we call OpenGenome2, was compiled from curated, non-redundant nucleotide sequence data”

direct quote[Results, 'Designing mammalian chromatin patterns']

“By contrast, when we repeated the same design pipeline except with a uniform or bigram proposal replacing Evo 2, we observed poorer token-matched inference-time scaling and poor consensus among ensemble predictions”

absence[Methods, 'Evo 2 architecture, training, and data']

No formal power analysis for dataset size or outlier handling is reported.

PASSReviewer 2· 88% conf

The study is a computational model development and evaluation. The relevant design elements—data curation, model architecture, training strategy, evaluation benchmarks—are described in adequate detail.

Evidence

direct quote[Methods, paragraph 1]

“As in Evo 1, we excluded genomic sequences from viruses that infect eukaryotic hosts from the training data for biosafety purposes.”

paraphrase[Methods, 'Evo 2 architecture, training, and data' section]

“The paper describes StripedHyena 2 architecture, two-phase training (pretraining at 8k context, midtraining extending to 1M), and numerous evaluation benchmarks.”

paraphrase[Results, various sections (Figures 2, 3, 4; Extended Data Figures)]

“The paper compares Evo 2 against multiple baselines (Evo 1, Nucleotide Transformer, ESM-2, GPN-MSA, PhyloP, etc.) across most evaluation tasks.”

PASSReviewer 3· 90% conf

The study design is well-defined for a computational foundation model: training data curation, architecture choices, multi-stage pretraining, and evaluations are described in adequate detail.

Evidence

direct quote[Results, Evo 2 architecture, training, and data, paragraph 2]

“As in Evo 1, we excluded genomic sequences from viruses that infect eukaryotic hosts from the training data for biosafety purposes.”

direct quote[Results, Human variant effect prediction, paragraph 1]

“We also contextualize the performance of Evo 2 against a wide range of models, including statistical measures of conservation (for example, PhyloP); unsupervised language models of proteins, RNA and DNA (for example, ESM-1b); supervised splicing prediction models (for example, Pangolin and SpliceAI); and human variant effect prediction models (for example, AlphaMissense, GPN-MSA and CADD).”

direct quote[Results, Evo 2 architecture, training, and data, paragraph 2]

“with data weighting focused on genic windows to learn functional genetic elements”

The paper is primarily computational, training on DNA sequences from diverse species. The experimental validation uses cell lines (mESCs, HEK293T, K562), but no individual subject metadata (sex, age, health) is applicable. The species and strains are not reported as biological variables because the focus is on model development, not a biological study of organisms.

Sub-criteria

demographicsN/A

sex reportedN/A

sex justifiedN/A

age weight healthN/A

housing conditionsN/A

species strain sourceN/A

Per reviewer (3) — the votes behind the verdict

N/AReviewer 1· 95% conf

The study does not involve biological subject-level variables; it uses genomic sequences and cell lines, but no animal or human subjects. All sub-criteria are not applicable.

PASSReviewer 2· 85% conf

The study primarily analyzes genomic sequence data, where the biological variables are the sequences themselves. The species and data sources are well-documented.

Evidence

direct quote[Abstract; Results, 'Evo 2 architecture, training, and data'; Figure 1]

“Evo 2, a biological foundation model trained on 9 trillion DNA base pairs from a highly curated genomic atlas spanning all domains of life”

direct quote[Results, 'Evo 2 architecture, training, and data', Figure 1b legend]

“Each point in the UMAP (uniform manifold approximation and projection) graph represents a single genome in the training dataset that is embedded on the basis of the genome's k-mer frequencies.”

N/AReviewer 3· 95% conf

The study is a computational model trained on genomic sequences; it does not involve any new biological experiments with subjects, cell lines, or organisms, so biological variables are not applicable.

Evidence

paraphrase[Methods (implicit in model description throughout)]

“The entire study revolves around training and evaluating a computational model on genomic sequence data; no new biological experiments involving subjects or cell lines are described.”

The study trains on the OpenGenome2 dataset, which is derived from public sequence databases. For the human variant predictions, it uses public databases (ClinVar, BRCA1 SGE data, SpliceVarDB). The chromatin accessibility experiments involve mouse and human cell lines (mESCs, HEK293T, K562). For cell line work, ethics approval is not typically required but the paper does not mention any IACUC or IRB statements. Given that this is primarily a computational study using cell lines for validation, the need for ethics approval is not explicitly applicable, though a statement might be expected. I will mark it not_applicable because the primary computational work does not involve subjects, and the cell line validation is a minor part.

Sub-criteria

iacuc statementN/A

informed consentN/A

irb ethics statementN/A

regulatory complianceN/A

Per reviewer (3) — the votes behind the verdict

WARNReviewer 1· 80% conf

The paper reports no ethics approval for the experimental validation involving mouse embryonic stem cells (mESCs) and human cell lines, which is a gap requiring attention.

Evidence

absence[Methods, 'Designing mammalian chromatin patterns']

No IACUC statement is present for mESC experiments.

absence[Methods, 'Designing mammalian chromatin patterns']

No ethics approval for human cell line work is reported.

N/AReviewer 2· 95% conf

The study uses only public, de-identified genomic sequence data and public datasets (ClinVar, SpliceVarDB). No new human or animal subject data were collected. The chromatin accessibility experiments used cell lines, but the paper does not state ethics approval for those experiments.

Evidence

direct quote[Abstract; Results, 'Evo 2 architecture, training, and data']

“Evo 2 was trained on prokaryotic and eukaryotic genetic sequences... This new training dataset, which we call OpenGenome2, was compiled from curated, non-redundant nucleotide sequence data”

paraphrase[Results, 'Human variant effect prediction']

“The paper uses annotations from ClinVar, SpliceVarDB, and BRCA1 saturation mutagenesis data, all of which are public datasets.”

N/AReviewer 3· 90% conf

The study uses only public, de-identified genomic datasets (OpenGenome2) and pre-existing benchmark data; it does not collect new human or animal subject data, so ethics approvals are not required.

Evidence

direct quote[Data availability]

“The OpenGenome2 dataset used to train Evo 2 is available at: https://huggingface.co/datasets/arcinstitute/opengenome2”

direct quote[Results, Human variant effect prediction, paragraph 1]

“Using the ClinVar database, we compared the ability of Evo 2 against other methods for predicting the pathogenic effects of human genetic variants”

The paper uses mouse ESCs, HEK293T, and K562 cells for experimental validation, but does not report authentication (e.g., STR profiling) or mycoplasma testing. The DNA synthesis reagent is mentioned only in acknowledgments (Ansa Biotechnologies) without catalog numbers. Software tools are thoroughly identified with version and URLs. Given that cell lines and reagents are used but inadequately described, the score is 'fail'.

Sub-criteria

mycoplasma testingNOT REPORTED

reagents identifiedINADEQUATE

organisms identifiedN/A

antibodies identifiedN/A

cell line authenticationNOT REPORTED

software tools identifiedADEQUATE

Per reviewer (3) — the votes behind the verdict

FAILReviewer 1· 85% conf

Key biological resources (cell lines) are not authenticated or tested for mycoplasma, and reagents are not fully identified with catalog numbers. Software is well-identified.

Evidence

absence[Methods, 'Designing mammalian chromatin patterns']

Cell lines used but no authentication reported.

direct quote[Acknowledgements]

“We thank Ansa Biotechnologies for providing synthesized and assembled DNA for our mESC Morse code experiments.”

direct quote[Code availability]

“Code and tools for model exploration are available at the following links: top-level code repository: https://github.com/arcinstitute/evo2”

PASSReviewer 2· 90% conf

The key resources (software, models, datasets) are comprehensively identified with URLs, DOIs, and version information. The paper uses computational resources, not biological reagents.

Evidence

direct quote[Code availability section]

“Code and tools for model exploration are available at the following links: top-level code repository: https://github.com/arcinstitute/evo2”

direct quote[Code availability section]

“Evo 2 40B: https://huggingface.co/arcinstitute/evo2_40b”

direct quote[Data availability section]

“The OpenGenome2 dataset used to train Evo 2 is available at: https://huggingface.co/datasets/arcinstitute/opengenome2”

WARNReviewer 3· 80% conf

Software tools and model parameters are well-identified and deposited, but the paper does not formally identify all key biological resources (e.g., antibodies, cell lines) used in the experimental validation portions.

Evidence

direct quote[Code availability]

“top-level code repository: https://github.com/arcinstitute/evo2”

paraphrase[Results, Designing mammalian chromatin patterns]

“The paper mentions using 'Bl6xcast mESCs', 'HEK293T' and 'K562' cells but does not report cell line authentication, source catalog numbers, or mycoplasma testing.”

direct quote[Results, Designing mammalian chromatin patterns, paragraph 1]

“We used an ensemble of Enformer and Borzoi to define a scoring function”

The paper reports exact p-values for key comparisons (e.g., Wilcoxon rank sum test, hypergeometric test) and provides confidence intervals where relevant (e.g., AUROC and AUPRC). Software used for analysis is implicitly Python/scipy. Data presentation includes scatter plots, box plots, and genome browser tracks with per-group n stated. The mathematical plausibility checks are generally satisfied as reported statistics are consistent with the data. The paper is transparent about model performance and limitations.

Sub-criteria

tests namedADEQUATE

exact p valuesADEQUATE

effect sizes ciADEQUATE

data presentationADEQUATE

software identifiedINADEQUATE

assumptions verifiedINADEQUATE

mathematical plausibilityPASS

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 85% conf

Statistical tests are named, exact p-values are reported, data are shown with individual points, and software is identified. However, assumptions verification and effect sizes with confidence intervals are lacking.

Evidence

direct quote[Results, 'Human variant effect prediction']

“P value calculated by two-sided Wilcoxon rank sum test.”

direct quote[Results, 'Human variant effect prediction']

“P value calculated by two-sided Wilcoxon rank sum test.”

absence[Methods]

Assumptions not verified.

PASSReviewer 2· 85% conf

The statistical analysis is generally appropriate and well-reported. Many comparisons use standard metrics (AUROC, AUPRC, Spearman correlation, p-values from Wilcoxon rank sum tests). The paper reports exact p-values for key comparisons and effect sizes (AUROC, AUPRC).

Evidence

direct quote[Results, 'Designing mammalian chromatin patterns']

“two-sided Welch's t-test, P = 3.6 × 10^{-7}”

direct quote[Results, 'Genome-scale generation' and 'Designing mammalian chromatin patterns']

“one-sided hypergeometric P = 2.0 × 10^{-4}”

direct quote[Results, 'Human variant effect prediction' and 'Evo 2 learns evolutionary constraint']

“AUROC = 0.66, AUPRC = 0.15”

PASSReviewer 3· 88% conf

Statistical analyses are appropriate for a computational benchmark study: tests are named, exact p-values and effect sizes are reported, and data presentation is thorough with individual points and defined error bars.

Evidence

direct quote[Results, Designing mammalian chromatin patterns, paragraph 3]

“two-sided Welch's t-test, P = 3.6 × 10^{-7}”

direct quote[Results, Human variant effect prediction, paragraph 3]

“Evo 2 40B (AUROC = 0.66, area under the precision-recall curve (AUPRC) = 0.15) outperformed other genomic language models (AUROC range 0.50–0.59, AUPRC range 0.09–0.12)”

paraphrase[Figures 2, 3, 5, 6, and accompanying legends]

“Figures display scatter plots, individual data points, and error bars defined as SD, SEM, or CI as noted in their legends.”

The OpenGenome2 dataset is on Hugging Face. All code (Evo 2 training, inference, Evo Designer, SAE visualization) is on GitHub. Model weights are on Hugging Face. ATAC-seq data for mESC and human cell line experiments are deposited in SRA under BioProject accession numbers. This is a model of open science.

Sub-criteria

code sharingADEQUATE

accession numbersADEQUATE

repository depositADEQUATE

data availability statementADEQUATE

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 95% conf

All data and code are publicly available in repositories with persistent identifiers, including the OpenGenome2 dataset on HuggingFace, SRA accession numbers for ATAC-seq data, and multiple GitHub repositories for code.

Evidence

direct quote[Data availability]

“The OpenGenome2 dataset used to train Evo 2 is available at: https://huggingface.co/datasets/arcinstitute/opengenome2”

direct quote[Data availability]

“Raw reads for ATAC-seq analysis of mESC chromatin accessibility designs have been uploaded to the Sequence Read Archive (SRA) under BioProject accession number PRJNA1314301”

direct quote[Code availability]

“Code and tools for model exploration are available at the following links: top-level code repository: https://github.com/arcinstitute/evo2”

PASSReviewer 2· 98% conf

The paper provides excellent data and code availability. The training data, model weights, code, and evaluation tools are all deposited in public repositories with persistent identifiers.

Evidence

direct quote[Data availability section]

“The OpenGenome2 dataset used to train Evo 2 is available at: https://huggingface.co/datasets/arcinstitute/opengenome2”

direct quote[Data availability section]

“Raw reads for ATAC-seq analysis of mESC chromatin accessibility designs have been uploaded to the Sequence Read Archive (SRA) under BioProject accession number PRJNA1314301”

direct quote[Code availability section]

“Code and tools for model exploration are available at... https://github.com/arcinstitute/evo2”

PASSReviewer 3· 95% conf

Data and code availability are exemplary: the training dataset (OpenGenome2), model parameters, training and inference code, and analysis tools are all publicly deposited with persistent identifiers.

Evidence

direct quote[Data availability]

“The OpenGenome2 dataset used to train Evo 2 is available at: https://huggingface.co/datasets/arcinstitute/opengenome2”

direct quote[Code availability]

“top-level code repository: https://github.com/arcinstitute/evo2”

direct quote[Data availability]

“Raw reads for ATAC-seq analysis of mESC chromatin accessibility designs have been uploaded to the Sequence Read Archive (SRA) under BioProject accession number PRJNA1314301”

The methods section is extensive, covering model architecture, training data, evaluation protocols, and experimental validation. Limitations are explicitly acknowledged (e.g., underperformance on viral sequences, need for experimental validation of designs). Conclusions are generally proportional to the evidence presented; the authors do not overclaim functional or clinical readiness. Funding sources and competing interests are fully disclosed. No trial registration or specific reporting guideline is applicable for this computational study.

Sub-criteria

funding coiADEQUATE

trial registrationN/A

reporting guidelineN/A

methods completenessADEQUATE

all outcomes reportedADEQUATE

limitations discussedADEQUATE

conclusions proportionalADEQUATE

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 85% conf

The paper reports most outcomes transparently, discusses limitations, draws proportional conclusions, and includes funding and competing interests. Methods are detailed for computational work but lack full protocols for wet-lab experiments.

Evidence

direct quote[Results, 'Genome-scale generation']

“these evaluation metrics do not guarantee functional or replication-competent genomes, and our genome-scale generations lack important elements, such as some essential genes.”

direct quote[Competing interests]

“Competing interests M.G.D. acknowledges outside interest in Stylus Medicine...”

absence[Methods, 'Designing mammalian chromatin patterns']

Methods for wet-lab experiments not fully detailed.

PASSReviewer 2· 90% conf

The reporting is thorough and transparent. Methods are detailed, limitations are discussed, conclusions are proportional, and funding/COI are disclosed.

Evidence

direct quote[Results, 'Genome-scale generation' and 'Feature interpretation in Evo 2']

“Of note, these evaluation metrics do not guarantee functional or replication-competent genomes, and our genome-scale generations lack important elements, such as some essential genes.”

direct quote[Results, 'Human variant effect prediction']

“Evo 2 is competitive with widely used ProGen language models for protein DMS and with RNA language models for ncRNA DMS, although it underperforms state-of-the-art models on protein DMS.”

direct quote[Acknowledgements and Competing interests sections]

“Acknowledgements... Competing interests M.G.D. acknowledges outside interest in Stylus Medicine...”

PASSReviewer 3· 92% conf

The manuscript is transparently reported: methods are detailed, limitations are discussed, conclusions are proportional, and funding/competing interests are disclosed.

Evidence

direct quote[Results, Genome-scale generation, paragraph 1]

“Consistent with poor performance for viruses that infect humans on the language modelling task and on function prediction downstreams, Evo 2 also has poor performance on generating proteins from human viruses.”

direct quote[Acknowledgements and Competing interests]

“Acknowledgements We thank... P.D.H. acknowledges funding support from Arc Institute, Yosemite, Rainwater Foundation... B.L.H. acknowledges funding support from Arc Institute, the Gates Foundation... M.G.D. acknowledges outside interest in Stylus Medicine...”

direct quote[Abstract and Discussion]

“These evaluation metrics do not guarantee functional or replication-competent genomes, and our genome-scale generations lack important elements, such as some essential genes.”

7 copyedit issues flagged: mostly typo, consistency, clarity.

Checked 59 references: 46 verified — 13 unresolved.

Reproducibility

18 data/code links checked; 15 live.

Αlpha¹ Rigor Action ItemsFinal pass before journal submission

As a post-publication audit, the paper is robust in its core computational methodology and data sharing, but an informed reader should weigh the missing cell line authentication, mycoplasma testing, and ethics approval statements as significant reporting gaps that would warrant a correction or at minimum a published comment. The unresolved references (13 not found in registry) also raise a concern that should be investigated independently.

1.
HIGHethics
Add an IACUC approval statement (protocol number and institution) for the mouse ESC experiments in the Methods, and an IRB or institutional ethics statement for human cell line use (HEK293T, K562).
Missing ethics approval for animal-derived and human cell lines is a reporting gap that journals typically require; without it, the published article is incomplete.
2.
HIGHrigor
Add cell line authentication (STR profiling) and mycoplasma testing statements for mESCs, HEK293T, and K562 cells in a new 'Cell line authentication' subsection of Methods.
Current best practices require authentication and mycoplasma testing for cell lines used in experiments; their absence undermines reproducibility.
3.
HIGHreporting
Verify the 13 references flagged as 'not_found_in_registry' (see citation verification list) and correct or remove any that cannot be confirmed as legitimate publications.
References that cannot be located in any registry may indicate fabrication; at minimum, they impede reader verification of cited work.
4.
HIGHrigor
Provide catalog numbers and vendors for DNA synthesis reagent (Ansa Biotechnologies) and for cell lines (e.g., HEK293T from ATCC CRL-3216) in the Methods section.
Incomplete identification of key resources hinders reproducibility; catalog numbers allow other researchers to obtain identical materials.
5.
HIGHrigor
Add an explicit outlier handling procedure for computational predictions (e.g., filtering low-confidence predictions) in the Methods.
Without stated outlier handling, readers cannot assess whether selective reporting affects the results.
6.
HIGHstatistics
Report effect sizes with 95% confidence intervals for key comparisons (e.g., AUROC/ AUPRC differences between models) in the Results.
Confidence intervals provide information on precision and are critical for interpreting the practical significance of differences.
7.
HIGHstatistics
Verify assumptions (normality, variance homogeneity) for the Welch's t-tests and Wilcoxon tests, or state why non-parametric tests were chosen without verification.
Readers cannot judge whether the reported p-values are accurate without knowing whether test assumptions were met.
8.
HIGHreporting
Detail wet-lab protocols (cell culture conditions, transfection reagents and concentrations, ATAC-seq library preparation) in the Methods or supplement.
Insufficient methods detail prevents independent replication of the experimental validation.
9.
MEDIUMcopyedit
Fix typo 'Distribut i on' to 'Distribution' in Results 'Genome-scale generation' and Extended Data Fig. 10f.
Typographical errors in published figures and text reduce perceived quality.
10.
MEDIUMcopyedit
Standardize hyphenation to 'rank-sum test' and fix tense inconsistency ('outperformed ... performs') in Results 'Human variant effect prediction'.
Consistency in terminology and grammar improves readability and professional appearance.
11.
MEDIUMreporting
Clarify in the Methods that Evo 2 is trained on human genome sequences (as part of multi-species data) but not on labeled human variant data, to avoid misinterpretation of the claim 'not trained on any human genetic variation'.
The current phrasing could mislead readers into thinking no human sequences were used; clarification prevents unwarranted claims of zero-shot capability.
12.
MEDIUMreporting
For the claim that Evo 2-generated mitochondrial sequences have 'the correct number of CDSs, tRNA genes, and rRNA genes expected in human mitochondria', add a caveat that this is an in silico estimate (MitoZ) and not experimental validation.
Overstating in silico predictions as 'correct' misrepresents the strength of evidence.
13.
LOWreporting
Report the number of independent biological and technical replicates for the ATAC-seq experiments in the chromatin accessibility validation.
Replicate information is essential for assessing the reliability of wet-lab results.
14.
LOWreporting
Add a supplementary table listing all evaluation datasets with sample sizes, exclusion criteria, and URLs for full traceability.
Centralizing dataset metadata aids reproducibility and reviewer/reader verification.

Full text (93,798 chars · analyzed ~80%, head + tail)Analyzed: Jul 5, 2026

Read alongside

Genome modelling and design across all domains of life with Evo 2.

Open DOI

DOI

10.1038/s41586-026-10176-5

Share this Verified Rigor Record

Post on X Share on LinkedIn Share via Email

Record fingerprint

018aa25e7d717ee5297d790bd5a0f989b4c61a986b17de6c4bb7327005bee146

Verify via API

Issued Jul 5, 2026Engine 7.5.0

Αlpha¹ assesses methodological rigor, not the importance of the findings. See how we calibrate →