{"id":"3ac5c7ad-2d49-4355-ac9a-69e4ef53b91a","arxiv_id":"2508.20130","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":0.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A narrative review of AI and explainable AI for CRISPR guide RNA design and off-target prediction, with no new experimental or computational findings.","lead":"This preprint is a literature review of artificial intelligence and explainable AI methods applied to CRISPR guide RNA design and off-target safety. It summarizes recent tools and clinical studies, but contains no new research results and is drafted with placeholder text and unfinished figures.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 1.1 asserts that AI and explainable AI aided sgRNA selection and underpinned clinical development of exa-cel and other CRISPR therapies, but the cited clinical papers contain no such claim; this unsupported attribution supports the review's central narrative.","rationale":"The reader's weakest_assumption correctly identifies the unsupported clinical AI attribution as the load-bearing premise. My independent reading of the manuscript confirms this. The paper is an incomplete review draft with placeholders and truncations, and its central claim is an opinion about XAI's value. The only concrete bridge from that opinion to clinical reality is the Section 1.1 assertion that AI tools aided sgRNA selection for exa-cel and other therapies, and that explainable AI underpinned development. This assertion is made without a citation and is contradicted by the cited clinical literature, which does not discuss AI-based guide selection. If the authors cannot substantiate this, the review's emphasis on explainable AI as a clinical enabler is not supported by the cited evidence. No other concern is more fundamental: the paper is not making a new mathematical or experimental claim, so the correctness risk is concentrated in this unsupported factual premise. The test I propose would settle the matter by direct literature verification. Since the reader's verdict is already UNVERDICTED due to the draft's incomplete state and lack of new research claims, my concern does not change that verdict; it reinforces it by adding a specific unresolved factual error risk.","tokens_in":22810,"tokens_out":2590,"duration_ms":26580,"concrete_test":"Check the methods and supplementary protocols of references [24] (Frangoul et al., NEJM 2021), [25]–[26] (CLIMB trials, NEJM 2024), [27] (Gillmore et al.), and [30] (Chiesa et al.) for any mention of machine learning, deep learning, AI, or explainable AI in the guide RNA selection process. Specifically, search the full texts and supplements for terms like 'sgRNA design', 'guide selection', 'in silico', 'algorithm', 'neural network', 'model', and 'explainable'. If no AI/XAI method is described or cited, the paper's claim in Section 1.1 is unsupported. As an additional positive check, look for any pre-registered computational design tool or companion bioinformatics paper for exa-cel and NTLA-2001; if such papers exist, they would need to be explicitly cited and would likely describe conventional rule-based or alignment-based tools rather than explainable deep learning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The review's central narrative is that explainable AI is a key enabler of CRISPR clinical safety and success. The only concrete clinical evidence offered for this is in Section 1.1: 'Modern AI tools further aided in selecting an optimal sgRNA with high on-target activity and low off-target propensity' for exa-cel, and the broader claim that 'explainable AI-driven design has underpinned' clinical development. This is load-bearing because if exa-cel and the other cited therapies did not actually use AI (let alone explainable AI) for sgRNA design, then the review's emphasis on XAI as a clinical enabler lacks a factual foundation. The cited references—Frangoul et al. (refs [24]–[26]) for exa-cel, Gillmore et al. [27], and Chiesa et al. [30]—report clinical outcomes and do not describe AI-based sgRNA selection in their methods. The paper provides no primary source for this assertion. This is not merely a matter of 'outside current consensus'; it is an internally unsupported factual claim used to justify the review's thesis. If the claim is false, the conclusion's appeal to clinical success as evidence of XAI's value collapses.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a narrative review of AI and explainable AI (XAI) for CRISPR guide RNA design and off-target safety. It surveys recent machine-learning models for on-target activity prediction (CRISPRon, Kim et al., AttCRISPR, CRISPRedict, etc.), interpretability methods (attention, SHAP, integrated gradients), off-target prediction and experimental detection (CHANGE-seq/DISCOVER-Seq, etc.), and clinical applications of CRISPR (exa-cel, NTLA-2001, EDIT-101, base-edited CAR7 T cells). The authors argue that XAI is becoming a key enabler of efficient, safe, and clinically viable CRISPR therapies, and they conclude with a normative statement that XAI is an 'empowering enabler' of genome engineering. The review contains a large reference list (2020--2024) and two summary tables, but also includes several placeholder citations '(Ref.docx)', truncated sentences, and at least one unsupported clinical attribution.","tokens_in":23130,"tokens_out":4029,"duration_ms":37985,"significance":"The review addresses an important and timely topic: the role of explainable AI in CRISPR guide design and safety. If properly revised, it could serve as a useful interdisciplinary introduction, especially for experimental biologists seeking orientation in the fast-growing AI/CRISPR literature. The paper usefully compiles recent models, tools, and clinical studies, and it explicitly acknowledges the need for experimental validation and for ethical/regulatory reflection. Its main limitation is that it is a broad narrative review with no systematic methodology, no critical benchmarking, and no quantitative comparison of the tools it discusses. The central thesis that XAI is a key clinical enabler rests partly on an unsupported historical claim about exa-cel's guide selection, and the current draft contains visible incompleteness artifacts. Therefore, the manuscript is not yet in publishable form, but the flaws are reparable within the scope of a review.","major_comments":[{"comment":"The sentence 'Modern AI tools further aided in selecting an optimal sgRNA with high on-target activity and low off-target propensity' and the later claim that 'explainable AI-driven design has underpinned' clinical development are load-bearing for the review's thesis. The cited references [24]–[26] (Frangoul et al. and Locatelli et al.) report clinical outcomes for exa-cel but do not describe AI-driven or explainable-AI-driven guide selection; no other primary source is given. This is an unsupported factual attribution, not a matter of interpretation. The authors should either provide a documented primary source showing that AI/XAI was used in guide selection for exa-cel, or explicitly reframe these statements as recommendations or as plausible future uses of XAI rather than as historical facts.","section":"Section 1.1 (Ex Vivo Hematopoietic Stem Cell Therapies)"},{"comment":"The concluding assertion—'The role of explainable AI in enhancing CRISPR precision and safety is that of an empowering enabler'—is presented as if it were a finding of the reviewed literature. The manuscript surveys examples of AI and XAI in gRNA design, but it does not provide comparative or causal evidence that XAI specifically improved clinical outcomes, safety, or regulatory decisions. Since this normative claim is the capstone of the review, it should be explicitly framed as the authors' perspective or as a research agenda, not as a conclusion derived from the cited studies.","section":"Conclusion (Section 4)"},{"comment":"The manuscript contains multiple placeholder citations '(Ref.docx)' and truncated sentences that indicate an incomplete draft. Examples include: '(Ref.docx) (Ref.docx)' after 'cleavage efficiency' in the Introduction; '(Ref.docx)' after 'with unprecedented sensitivity' in the Off-target Prediction section; the broken sentence 'a model might predict 90Explainable AI techniques thus enrich...' in the XAI section; and 'they found that around 6Another layer of complexity...' in the Off-target section. These artifacts prevent a reader from evaluating the completeness and accuracy of the review. They must be fixed before the manuscript can be considered further.","section":"Throughout: Introduction, AI Models, Off-target Prediction"},{"comment":"The text repeatedly refers to the experimental method as 'CHANGE-seq', but reference [7] is titled 'COCHANGE-seq'. This is a factual inconsistency in a central methodological term. (The corresponding Nature Biotechnology article is normally cited as 'CHANGE-seq'; the 'CO' prefix may be an error in the reference list.) The authors should verify the correct assay name and ensure that text, citations, Table/Figure captions, and the bibliography are consistent.","section":"Section 3 (Off-target Prediction) and Reference [7]"}],"minor_comments":[{"comment":"The parenthetical '(2020–2025, updated to reflect current year if needed)' is a template artifact and should be removed or replaced with the actual coverage dates.","section":"Abstract"},{"comment":"The keyword list contains 'Nature Biology', which appears to be a stray phrase rather than a keyword. It should be deleted.","section":"Keywords"},{"comment":"All four figures are labeled '(Conceptual figure placeholder)'. If these are not final figures, they should be either implemented as proper, informative schematic figures or removed; placeholders are not appropriate for a journal submission.","section":"Figures"},{"comment":"There are typographical errors in the text: 'nowexagamglogene' should be 'now exagamglogene'; 'editit' appears in the AI Models section; 'AA V' is written with a space in several places; 'H”oijer' has a mangled quote character in the Off-target section; 'P ASTEsystem' should be 'PASTE system'. A careful proofreading pass is needed.","section":"Section 1.1"},{"comment":"Earlier sections (AI Models for gRNA Design, Explainable AI Techniques, Off-target Prediction and Safety) are unnumbered, while '1 Clinical Applications' begins numbered sections. The numbering scheme should be made consistent throughout.","section":"Section numbering"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a broad review, so the lack of original benchmarking is not by itself disqualifying. However, the draft appears to have been submitted in a partially assembled state, with placeholder citations, truncated sentences, and an unsupported historical claim about AI's role in exa-cel's design. The central thesis can be defended if the exa-cel attribution is either sourced or softened, but in its current form the review is not ready for a serious journal. I recommend major revision, with scrutiny of the factual claims about clinical trial methodology."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: this is an unfinished survey, not a research paper, and its central narrative rests on an unsupported claim that AI and explainable AI underpin clinical CRISPR successes like exa-cel. The cited clinical papers don't say that. So the review's thesis overreaches its evidence.\n\nWhat the paper does well: it brings together a reasonable list of AI models for guide design — CRISPRon, AttCRISPR, CRISPRedict, the BERT off-target model, R-CRISPR — and it sketches the off-target detection landscape (GUIDE-seq, DISCOVER-Seq, etc.) accurately enough for a newcomer. The sections on base editing, prime editing, and PASTE are up to date and cite the right primary sources. As a broad orientation to the literature, it has some value.\n\nThe soft spots are not minor. Section 1.1 asserts that 'modern AI tools further aided in selecting an optimal sgRNA' for exa-cel and that 'explainable AI-driven design has underpinned' clinical development. The cited Frangoul, Locatelli, and Gillmore papers describe no AI-based guide selection. This is load-bearing because the conclusion's appeal to clinical success as validation of XAI collapses without it. The paper also has placeholders like '(Ref.docx)', conceptual figure placeholders, a truncated sentence, and an author name spelled two ways on the title page. There's a method-name error: the text says CHANGE-seq but reference [7] is COCHANGE-seq. These are fixable, but they signal the draft is unfinished.\n\nOn the whole, the manuscript is not ready for peer review. The general claim that AI improves gRNA design is true, but the review's specific contribution — that explainable AI is a key clinical enabler — is not supported. I would not cite this in its current form, and I wouldn't bring it to a reading group except as an example of why unsupported attributions matter. If the authors are willing to revise, correct the clinical claims, and complete the draft, it could become a useful review for graduate students entering the field. As it stands, I'd recommend the editor return it without external review.\n\nBest,\n[Your name]","headline":"Unfinished review with a load-bearing unsupported claim linking explainable AI to clinical CRISPR successes; not publishable in current form.","tokens_in":23526,"tokens_out":3516,"would_cite":false,"duration_ms":30970,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A review argues that explainable AI can open the black box of CRISPR guide design and sharpen off-target safety.","keywords":["CRISPR","guide RNA design","deep learning","explainable AI","off-target prediction","genome editing safety","base editing","clinical translation"],"falsifier":"Inspect the methods or supplementary files of the exa-cel trials [24], the NTLA-2001 trial [27], and the EDIT-101 trial [31]: if the guide sequences are justified by standard rule-based or expert selection without any machine-learning or explainable-AI step, the paper's assertion that modern AI tools aided sgRNA selection is contradicted.","tokens_in":22751,"feed_emoji":"🧬","tokens_out":4411,"duration_ms":42851,"temperature":0.7,"pith_summary":"This review assembles evidence from 2020–2025 that modern machine learning, especially deep learning, now outperforms older rule-based methods at predicting how efficiently a guide RNA cuts its intended site and how likely it is to cut elsewhere. The authors' central thesis is that explainable AI techniques—attention weights, SHAP values, and interpretable linear models—are what make these powerful black boxes usable in high-stakes genome editing, because they turn a single prediction score into a sequence-position map that matches known biology and can be audited. If true, the practical consequence is that guide selection can move from expensive trial-and-error to in silico screening followed by targeted validation, with a more defensible off-target safety case for regulators.","feed_headline":"Explainable AI can demystify CRISPR guide design","feed_subtitle":"A review finds attention maps and SHAP values turn black-box gRNA predictions into decisions a clinician can audit.","key_machinery":"The central object is the guide RNA's 20-nucleotide targeting sequence treated as a sequence-prediction input. The machinery that carries the argument is a pipeline of deep learning architectures—CNNs for motif detection, recurrent and attention layers for positional dependencies, and BERT-style pre-training for guide-target pairs—each equipped with an attribution method (attention weights, SHAP, integrated gradients) that projects the model's decision back onto individual nucleotide positions. Genome-wide off-target assays such as CHANGE-seq and DISCOVER-Seq+ supply the empirical ground truth on which both the predictive models and the explainability claims rest.","core_discovery":"The paper's discovery, in its own terms, is a convergence: deep learning models like CRISPRon and CRISPR-Net show that sequence context plus epigenetic data materially improve gRNA efficiency forecasts, while interpretability studies such as AttCRISPR and a BERT-based off-target model show that attention and SHAP analyses can reveal biologically meaningful position sensitivities—for instance, PAM-proximal seed-region positions driving cleavage. On the safety side, the review contends that off-target prediction has shifted from mismatch-counting rules to learned models trained on genome-wide empirical maps, and that these predictions are credible enough to shape clinical guide selection, espe","pith_inferences":["The paper treats explainability as a trust-building feature, but it implies a stronger hypothesis: that interpretable models will discover genuinely new sequence determinants of Cas activity, not merely rediscover known seed-region and PAM effects. That hypothesis is testable by prospectively validating a novel motif flagged by attribution analysis.","The clinical-translation narrative would be directly testable by auditing how the guides for exa-cel, NTLA-2001, and EDIT-101 were actually chosen; the cited clinical reports are not explicit about AI involvement, so the review's claim that AI underpins those successes should be read as an aspiration until such an audit is performed.","The same attention/SHAP machinery can be extended to base editors and prime editors, attributing their distinct byproducts—off-target base conversions, RNA edits, pegRNA-related indels—to sequence context, a measurable extension the paper does not carry out."],"forward_implications":["Guide design shifts from wet-lab screening of many candidates to in silico pre-screening, with explainability telling experimentalists which positions to verify first.","Off-target safety becomes a per-patient question: variant-aware models can flag guides whose risk profile changes with a person's SNPs, making population-aware guide panels feasible.","Newer editors—base editors, prime editors, and integrase-based PASTE—need the same explanatory machinery for outcomes that no longer follow the classic double-strand-break off-target pattern.","Jointly modeling on-target potency and off-target risk lets designers optimize the trade-off explicitly rather than applying a fixed mismatch threshold.","Explainable reasons, such as a mismatch at a position learned to be critical for Cas9, can be packaged into regulatory and clinical risk documentation for approved therapies."],"supporting_citations":[{"why":"Supplies the deep-learning on-target predictor CRISPRon, which integrates sequence and epigenetic features to improve gRNA efficiency prediction.","marker":"[1]"},{"why":"Supplies the BERT-based off-target model that uses attention and SHAP analysis for interpretability, anchoring the explainable-AI section.","marker":"[4]"},{"why":"Supplies a deep-learning off-target predictor with a novel sgRNA-DNA encoding, supporting the shift from mismatch-counting to learned models.","marker":"[5]"},{"why":"Supplies the large-scale CHANGE-seq off-target dataset that maps where Cas9 actually cuts, providing empirical ground truth for predictive models.","marker":"[7]"},{"why":"Supplies the population-variant off-target predictor, grounding the claim that SNPs can create or abolish off-target sites.","marker":"[9]"},{"why":"Supplies CRISPRedict, an interpretable linear-model counterpoint showing that simpler models can match deep learning for efficiency prediction.","marker":"[16]"},{"why":"Supplies PASTE, the drag-and-drop integrase method that avoids double-strand breaks, grounding the safety-by-design discussion.","marker":"[23]"},{"why":"Supplies the exa-cel clinical trial, the main clinical anchor for the review's claim that carefully selected guides achieved therapeutic success.","marker":"[24]"},{"why":"Supplies the NTLA-2001 in vivo CRISPR trial, the key example of direct in-patient guide selection targeted to the TTR gene.","marker":"[27]"}],"fun_headline_variants":["AI explains CRISPR guide design for safer genome editing","Explainable AI advances CRISPR guide RNA safety","How AI is demystifying CRISPR guide RNA design","Transparent AI models improve CRISPR off-target prediction"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The paper's strongest claim depends on believing that the approved clinical CRISPR therapies it cites, especially exa-cel, actually relied on AI-based guide selection when their guides were chosen; the cited trial reports do not describe such AI, so if that premise fails the review loses much of its 'explainable AI underpins clinical success' narrative.","fun_headline_variants_meta":{"raw":{"variants":["AI explains CRISPR guide design for safer genome editing","Explainable AI advances CRISPR guide RNA safety","How AI is demystifying CRISPR guide RNA design","Transparent AI models improve CRISPR off-target prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000678,"raw_usage":{"total_tokens":2879,"prompt_tokens":666,"completion_tokens":2213,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":410,"completion_tokens_details":{"reasoning_tokens":2154}},"tokens_in":410,"tokens_out":2213,"duration_ms":16674,"temperature":1.0,"reasoning_tokens":2154,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:00:23.309136+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the methods or supplementary files of the exa-cel trials [24], the NTLA-2001 trial [27], and the EDIT-101 trial [31]: if the guide sequences are justified by standard rule-based or expert selection without any machine-learning or explainable-AI step, the paper's assertion that modern AI tools aided sgRNA selection is contradicted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the BERT-based off-target model that uses attention and SHAP analysis for interpretability, anchoring the explainable-AI section."},{"cited_title":"Bioinformatics 37(15), 2299–2307 (2021)","cited_arxiv_id":null,"evidence_quote":"Supplies a deep-learning off-target predictor with a novel sgRNA-DNA encoding, supporting the shift from mismatch-counting to learned models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the large-scale CHANGE-seq off-target dataset that maps where Cas9 actually cuts, providing empirical ground truth for predictive models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the population-variant off-target predictor, grounding the claim that SNPs can create or abolish off-target sites."},{"cited_title":"Nucleic Acids Res","cited_arxiv_id":null,"evidence_quote":"Supplies CRISPRedict, an interpretable linear-model counterpoint showing that simpler models can match deep learning for efficiency prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies PASTE, the drag-and-drop integrase method that avoids double-strand breaks, grounding the safety-by-design discussion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the exa-cel clinical trial, the main clinical anchor for the review's claim that carefully selected guides achieved therapeutic success."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the NTLA-2001 in vivo CRISPR trial, the key example of direct in-patient guide selection targeted to the TTR gene."}],"review_version":1}