{"id":"ac1070a3-ab64-4d00-9c7c-3726b3fda418","arxiv_id":"2412.12668","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":1.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A literature review that maps AI and deep learning methods for central-dogma-centric multi-omics integration and disease modeling.","lead":"This paper reviews how artificial intelligence and deep learning are being applied to combine DNA, RNA, and protein data in multi-omics disease research. A generalist might read it as a map of current methods, model families, and foundation models, though it offers no new experiments.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The review's evidence base is corrupted by multiple reference-mapping errors, so its field-level claim about AI multi-omics cannot be verified as written.","rationale":"The reader's weakest assumption is exactly the one I find least secure, and the full text supplies further violations beyond the two examples in the reader's report. I considered the abstract's unsupported clause about genetic-locus identification, but the citation-mapping failure is more damaging because it affects every summarized result, not just one omitted topic. A CONDITIONAL verdict is appropriate: the errors are fixable, but until the citation audit passes, the paper should not be used as a definitive reference. Hence no change to the reader's verdict.","tokens_in":18330,"tokens_out":6425,"duration_ms":59021,"concrete_test":"Write a small script to extract every [n] citation from the text and Tables 1-4 and compare each cited title in the reference list with the model or tool name in the citing sentence, and also check that each reference number is used for only one distinct model across the tables. If any mismatch remains after correcting the known cases, the survey's citation base is unreliable and the central claim should be treated as unsupported until a full citation audit is done.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that numerous AI multi-omics studies exist and improve disease prediction. Because the paper contributes no new results, this claim rests entirely on faithful attribution of more than 130 cited papers. That attribution is not faithful. Section 5.5 attributes CELLama and SCimilarity to [47] and [48], but the reference list assigns those numbers to DRPBind and SPHINKS. Table 4 lists CLCluster as [85], while [85] is NetICS and the sentence about CLCluster in Section 3.4 cites [101]; the same table lists GRMEC-SC as [101], although [101] is CLCluster and the GRMEC-SC paper is [100]. Eq. (4) also uses undefined Gr and sigma_r, making the Q-Former description incomplete. These are not isolated typos: they occur in the very structures, Tables 1-4, that carry the survey's evidence. When model names, reference numbers, and reference titles disagree, a reader cannot tell which reported accuracy supports which claim, so the central claim is not currently verifiable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey of artificial intelligence and deep learning methods for central-dogma-centric multi-omics data. It organizes more than 130 references into four task categories—classification, regression, generation, and clustering—and discusses foundation models, long-sequence modeling, evaluation benchmarks, and emerging frontiers. The paper's central claim, stated in the abstract, is that AI-driven multi-omics models have improved disease-prediction accuracy and advanced precision medicine. As a review, the paper contributes no new quantitative results; its value depends entirely on the accuracy of its synthesis, its mathematical descriptions, and its citation attributions.","tokens_in":18461,"tokens_out":6376,"duration_ms":51692,"significance":"If the reference attributions and mathematical descriptions were accurate, this survey would be a useful entry point for computational biologists: it covers a timely topic, organizes a large literature, and makes a plausible case that AI-based integration of genomic, transcriptomic, and proteomic data can improve disease modeling. The organization by task type and the emphasis on foundation models are strengths. However, the paper currently contains several concrete citation mismatches and one undefined equation, and because these occur in the tables and sections that carry the survey's evidence, the central claim cannot be fully verified as written. The appropriate standard for a survey is the reliability of its synthesis; that standard is not yet met.","major_comments":[{"comment":"The Q-Former paragraph introduces a formula with symbols G_r and sigma_r that are never defined, and the displayed expression does not correspond to the described 'learnable query embeddings interact with input features via multi-head attention and cross-attention' mechanism. Since the abstract promises mathematical definitions, an undefined equation is a substantive gap; either define all symbols and connect the formula to the Q-Former architecture, or delete it.","section":"Section 2, Eq. (4)"},{"comment":"The text states that CELLama [47] transforms cell data into 'sentences' and that SCimilarity [48] uses metric learning. In the reference list, [47] is DRPBind (Sharma et al., bioRxiv 2023) and [48] is SPHINKS (Migliozzi et al., Nature Cancer 2023); neither is the cited method. The actual CELLama and SCimilarity papers appear to be missing from the reference list, so this passage cannot be checked.","section":"Section 5.5, references [47] and [48]"},{"comment":"Table 4 assigns CLCluster to reference [85], but [85] is NetICS (Dimitrakopoulos et al., Bioinformatics 2018); the body text in Section 3.4 correctly cites CLCluster as [101]. The same table assigns GRMEC-SC to [101], but [101] is CLCluster and the GRMEC-SC paper is [100]; the scMDC row is listed as [100] although [100] is GRMEC-SC and scMDC is reference [103]. These mismatches, in the very table that carries the clustering evidence, make it impossible to verify which reported result belongs to which method.","section":"Table 4 and Section 3.4"},{"comment":"The abstract says the paper 'reviews the mathematical definitions of multi-omics,' but Section 2 only provides generic formulas for linear projection, an MLP, attention, and an undefined Q-Former expression; it never defines a multi-omics dataset or the integration problem mathematically. This discrepancy should be fixed either by adding a formal problem statement or by softening the abstract's claim.","section":"Abstract and Section 2"}],"minor_comments":[{"comment":"The opening sentence claims SVM was 'widely used in early multi-omics classification tasks' and cites [47,48]; however, [47] is a single-sequence binding-residue prediction method rather than a multi-omics integration study, and [48] is a single multi-omics subtype-classification paper. A broader and more apt set of citations is needed to support the 'widely used' claim.","section":"Section 3.1"},{"comment":"References [15] and [27] are the same article (Nemeth et al., Nature Reviews Genetics 2023) but are cited as separate entries; these should be unified into a single reference.","section":"References [15] and [27]"},{"comment":"Table 2 lists the venue for Seal et al. as 'Genomics 2022', but reference [59] is dated 2020 (Genomics 112(4), 2833-2841); the year should be aligned with the reference list.","section":"Table 2, row for Seal et al."},{"comment":"The text says 'Rosen et al. [127] developed a universal cell embedding,' but reference [127] is by Heimberg et al., not Rosen et al.; the author attribution is incorrect.","section":"Section 5.5, reference [127]"},{"comment":"Figure 1's timeline cites [35]-[40] as milestones for SVM, random forest, RNN, VAE, Transformer, and LLM, but these references are bioinformatics applications rather than the original sources of those methods; the figure should clarify that the timeline marks their emergence in multi-omics, not their invention.","section":"Figure 1"},{"comment":"The reference list has several formatting inconsistencies: many entries use 'al.' instead of 'et al.' and some author fields are malformed (e.g., [31] reads 'Huo, L., Jiao, J. Li, Chen, L.'); a full proofread of the reference list is needed.","section":"Reference list"}],"recommendation":"major_revision","confidential_remarks":"The paper's breadth is a strength, but the citation mismatches in Tables 1-4 and Section 5.5, together with the undefined Eq. (4), currently undermine the survey's reliability. I would ask the authors to re-verify every reference number in the tables and text, correct the missing/incorrect attributions, and either define or remove Eq. (4). After such a revision, a re-review focusing on citation accuracy would be appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a review, not a research paper, and it's a decent roadmap with a genuinely useful organizing idea — central dogma as the spine for AI multi-omics integration. But the reference layer is not trustworthy as printed, and until the citation mismatches are fixed, I wouldn't hand it to a student as a map.\n\nWhat it does well: the paper structures a large literature (130+ papers) by task type — classification, regression, generation, clustering — and then by architecture, with tables giving model, category, data, venue, and year. The central dogma framing is a reasonable way to organize DNA/RNA/protein integration work, and the discussion of foundation models (Evo, scGPT, LucaOne, CD-GPT) is current and sensible. For a computational biologist entering this area, the bird's-eye view is genuinely useful.\n\nSoft spots: the load-bearing part of a survey is faithful attribution, and that part is not reliable. Concretely: Section 5.5 cites CELLama and SCimilarity as [47] and [48], but those references are DRPBind and SPHINKS. Table 4 lists CLCluster as [85] when [85] is NetICS, and lists GRMEC-SC as [101] when [101] is CLCluster; the running text in Section 3.4 actually gets these right, so the tables and text disagree. Eq. (4) introduces Gr and sigma_r with no definition and doesn't obviously describe the Q-Former mechanism it's supposed to illustrate. These are localized, but they sit in exactly the tables and equations that carry a survey's evidentiary weight. If a reader can't trust the number-to-paper mapping, the reported accuracies become unverifiable. The central claim — that AI multi-omics models improve disease prediction — is plausible and supported by the literature generally, but this manuscript doesn't presently make that case cleanly.\n\nAlso worth noting: the text assigns [47] and [48] to SVM use in classification, which is consistent with DRPBind and SPHINKS, so the later misuse in Section 5.5 looks like a copy-paste or renumbering slip rather than a systematic problem. That makes me think the fixes are mechanical, not deep. But mechanical or not, they need to happen before this is used as a reference.\n\nBottom line: the survey deserves a serious referee — the field needs a good map like this, and the errors are correctable. I'd send it to review with instructions to verify every table entry and equation. For my own work, I wouldn't cite it until a corrected version appears.","headline":"A useful but currently unreliable survey: the central-dogma framing and task-based tables are good, yet citation mismatches and an undefined equation undermine its value as a reference until fixed.","tokens_in":18972,"tokens_out":2619,"would_cite":false,"duration_ms":22048,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This review argues that AI-driven multi-omics models—integrating DNA, RNA, and protein data along the central dogma—have improved disease prediction, helped identify disease-associated loci, and are advancing precision medicine.","keywords":["multi-omics","central dogma","artificial intelligence","machine learning","deep learning","foundation model","disease prediction","precision medicine"],"falsifier":"Re-run the headline benchmark comparisons on their stated public datasets and check whether the reported numbers reproduce; for example, test whether the 99.8% pan-cancer accuracy, the 99.16% five-dataset average, and the 94.68% graph-attention accuracy hold, and verify whether the references for CELLama and SCimilarity in Section 5.5 are the papers actually cited there. Any large shortfall or citation mismatch would undercut the review's map of the field.","tokens_in":18122,"feed_emoji":"🧬","tokens_out":11009,"duration_ms":90647,"temperature":0.7,"pith_summary":"This paper is a wide survey, drawing on more than 130 studies, of how artificial intelligence is used to analyze multi-omics data. Its central argument is that single-omics approaches—genomics, transcriptomics, proteomics, metabolomics taken alone—are too noisy and high-dimensional to reliably distinguish disease subtypes, whereas AI models that integrate multiple omics layers along the central dogma (DNA → RNA → protein) improve disease prediction and reveal disease-associated genetic loci. The review organizes the field into three fusion strategies, four deep-learning task families, and a new wave of foundation models pretrained on large-scale biological sequences. The practical payoff, if the surveyed results are accurate, is a pathway toward precision medicine in which disease subtypes and drug responses are predicted from integrated molecular profiles.","feed_headline":"AI models that fuse DNA, RNA, and protein sharpen disease prediction","feed_subtitle":"A survey of 130+ studies finds deep learning that integrates omics layers is pushing disease subtyping toward precision medicine.","key_machinery":"The conceptual axis is the central dogma of molecular biology ($\\mathrm{DNA} \\to \\mathrm{RNA} \\to \\mathrm{protein}$), which supplies the biological rationale for combining omics layers. The technical machinery is a short list of multimodal alignment mechanisms: the linear projector $Y = WX + B$, the multilayer perceptron with nonlinear activations $Y = f(W_2 f(W_1 X + B_1) + B_2)$, cross-attention $\\mathrm{Attention}(K,Q,V) = \\mathrm{softmax}(QK^{\\mathsf{T}} / \\sqrt{d_k})V$, and the Q-Former, a lightweight Transformer that converts variable-length inputs into fixed-length learned queries. The same section also defines the two-stage pretraining paradigm for foundation models—first on single-molecule sequences, then on multiple molecules with learned weights—as the mechanism for teaching models the cross-molecular interactions of the central dogma.","core_discovery":"The paper claims that deep learning has become the most effective tool for integrating the molecular layers described by the central dogma, and that this integration measurably improves disease modeling. It assembles a series of reported results: a self-normalizing network classifies 33 cancer types with 99.8% accuracy; a supervised multi-head attention transformer reaches an average 99.16% accuracy across five TCGA cancer datasets; a graph attention model improves cancer classification by about 4.69% over graph convolutional networks; an autoencoder improves liver-cancer survival prediction (C-index 0.68 versus 0.62 for PCA); regression models predict gene expression and drug response from DNA methylation, RNA, and protein features; and generative models produce synthetic single-cell and EHR data. For foundation models, the review points to large pretrained systems—cited examples include scGPT, Evo, CD-GPT, and LucaOne—that transfer to cell annotation, batch integration, perturbation response prediction, and gene-network inference. The organizing contribution of the review is a taxonomy: post-fusion strategies (linear projector, MLP, cross-attention, Q-Former) and four task families (classification, regression, generation, clustering), with central-dogma-aware two-stage pretraining as the emerging frontier.","pith_inferences":["The review's retrospective evidence does not itself prove that AI-driven multi-omics models improve disease prediction; the decisive test would be prospective: use such a model to nominate new disease loci or drug targets and validate them experimentally.","The taxonomy implies a controlled experiment the paper does not run: fix the downstream task and vary only the fusion strategy (linear projector versus MLP versus cross-attention) to separate the effect of fusion choice from model scale and data volume.","If the central dogma is a genuine biological prior, then cross-species transfer of foundation models should improve with model size; this is testable on the evaluation benchmarks the paper cites, extending the single-cell cell-atlas results to proteomic and metabolomic layers."],"forward_implications":["If the central claim holds, multi-omics deep-learning classifiers are already accurate enough (above 99% in some reported cases) that remaining progress in disease subtyping will come from data coverage and model interpretability rather than from any single omics layer.","Generative models become a practical solution for data sparsity and privacy: synthetic multi-omics and single-cell samples can be produced when real samples are scarce or too sensitive to share.","Foundation models pretrained on large molecular-sequence corpora will keep absorbing central-dogma structure, making transfer across cell types, tissues, and species more reliable.","Cross-attention and Q-Former-style fusion should spread from vision-language and RNA-generation tasks to general multi-omics integration, since they explicitly model interactions between DNA, RNA, and protein rather than concatenating features."],"supporting_citations":[{"why":"The Transformer paper that defines the attention mechanism reused as cross-attention in Equation (3).","marker":"[45]"},{"why":"The BLIP-2 Q-Former, presented as a lightweight cross-modal fusion strategy now applied to multi-omics integration.","marker":"[46]"},{"why":"The LASSO-MOGAT graph attention model, whose 94.68% TCGA classification accuracy supports graph-based multi-omics classification.","marker":"[51]"},{"why":"The supervised multi-head attention model whose 99.16% average accuracy across five TCGA cancer datasets supports transformer-based classification.","marker":"[54]"},{"why":"The self-normalizing model whose 99.8% pan-cancer accuracy supports the claim that deep networks excel at multi-omics classification.","marker":"[55]"},{"why":"The MOMA RNN framework, used as the main example of multi-omics regression with dimensionality reduction.","marker":"[37]"},{"why":"The CD-GPT two-stage pretraining approach that integrates central-dogma molecular interactions into a foundation model.","marker":"[65]"},{"why":"The Evo model, cited as the first foundation model for long-sequence multi-omics tasks.","marker":"[107]"},{"why":"The scGPT model pretrained on 33 million single cells, cited for cell annotation, multi-omics integration, and perturbation prediction.","marker":"[108]"},{"why":"The LucaOne foundation model, whose DNA-protein classification task directly evaluates whether a model captures cross-modal central-dogma interactions.","marker":"[109]"}],"fun_headline_variants":["AI fuses DNA, RNA, protein to boost disease prediction","Deep learning merges omics layers for sharper disease subtyping","99.8% cancer type accuracy from AI's central dogma fusion","Foundation model era: AI unifies central dogma data for diagnosis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The review is a secondhand report, so its conclusions stand or fall on whether the performance numbers and reference attributions it transcribes from the cited studies are accurate; in Section 5.5 the markers [47] and [48] are attached to CELLama and SCimilarity even though those references are DRPBind and SPHINKS, and Equation (4) contains undefined symbols.","fun_headline_variants_meta":{"raw":{"variants":["AI fuses DNA, RNA, protein to boost disease prediction","Deep learning merges omics layers for sharper disease subtyping","99.8% cancer type accuracy from AI's central dogma fusion","Foundation model era: AI unifies central dogma data for diagnosis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00088,"raw_usage":{"total_tokens":3826,"prompt_tokens":993,"completion_tokens":2833,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":2760}},"tokens_in":609,"tokens_out":2833,"duration_ms":19275,"temperature":1.0,"reasoning_tokens":2760,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:50:35.203902+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the headline benchmark comparisons on their stated public datasets and check whether the reported numbers reproduce; for example, test whether the 99.8% pan-cancer accuracy, the 99.16% five-dataset average, and the 94.68% graph-attention accuracy hold, and verify whether the references for CELLama and SCimilarity in Section 5.5 are the papers actually cited there. Any large shortfall or citation mismatch would undercut the review's map of the field.","supporting_citations":[{"cited_title":"bioRxiv, 2024– 0624600337 (2024)","cited_arxiv_id":null,"evidence_quote":"The CD-GPT two-stage pretraining approach that integrates central-dogma molecular interactions into a foundation model."},{"cited_title":"bioRxiv, 2024–0227582234 (2024)","cited_arxiv_id":null,"evidence_quote":"The Evo model, cited as the first foundation model for long-sequence multi-omics tasks."},{"cited_title":"Nature Methods, 1–11 (2024)","cited_arxiv_id":null,"evidence_quote":"The scGPT model pretrained on 33 million single cells, cited for cell annotation, multi-omics integration, and perturbation prediction."},{"cited_title":"bioRxiv, 2024–0510592927 (2024)","cited_arxiv_id":null,"evidence_quote":"The LucaOne foundation model, whose DNA-protein classification task directly evaluates whether a model captures cross-modal central-dogma interactions."}],"review_version":1}