{"id":"e18899c4-a046-490a-b944-0501630978f9","arxiv_id":"2505.09883","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DeepPlantCRE, a Transformer-CNN hybrid model, predicts plant gene expression levels from DNA sequence with reported accuracy up to 92.3% across multiple crop species.","lead":"This paper describes DeepPlantCRE, a deep learning model that combines a Transformer layer with convolutional neural networks to predict plant gene expression levels directly from promoter and surrounding DNA sequence. The authors report that the model outperforms two existing tools, DeepCRE and PhytoExpr, across five plant species and eight cotton genomes, and that motifs it learns match known transcription factor binding sites.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-species 92.3% rests on a label-construction step whose leakage status is unverified: quantile cutoffs appear to be computed on the full species before the chromosome-level split, and target-species labels are relative rather than absolute.","rationale":"The reader's weakest assumption is the same one I would stress: the within-species k-fold evaluation may leak label information because expression cutoffs are defined per dataset before the chromosome-level split. My attack sharpens this by pointing out the cross-species consequence: target-species labels are relative to each species' own percentiles, so the 92.30% transfer number is not anchored to absolute expression and has no locatable source in the text or tables. This is a correctness risk, not a stylistic complaint. It can be settled by inspecting and rerunning the provided code, which is a reasonable condition for acceptance. Because the paper's contribution is empirical and the reported numbers are plausible once the evaluation is made leak-free, I do not recommend moving the verdict from CONDITIONAL; the same conditional stance remains appropriate. I agree with the reader that the weakest point is the data-processing and cross-validation boundary, and I would not escalate to rejection unless the repository check reveals an unfixable leak.","tokens_in":12947,"tokens_out":10718,"duration_ms":115657,"concrete_test":"In the anonymous repository, run the preprocessing script that assigns low/medium/high labels and inspect whether the quantile thresholds are computed before the chromosome-level train/test split or inside each fold. Then rerun the full pipeline with thresholds recomputed on training chromosomes only and, for cross-species transfer, using the training species' thresholds to label the target species. If Tables 1 and 2 shift by more than a few points, or if the 92.30% cell in Figure 3 disappears, the central claim is inflated by leakage; if metrics are stable, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that DeepPlantCRE reaches 92.3% accuracy in cross-species transfer—requires a leak-free evaluation. Section 4.1 assigns low/medium/high labels from the 25th/75th percentiles of logMaxTPM \"within each dataset,\" and Section 4.2 then splits each species into k folds with k equal to the number of chromosomes. The text never states that the percentile thresholds are recomputed on the training chromosomes only. If the thresholds are estimated before the split, the label of every gene on a held-out chromosome is determined using expression values from training chromosomes; this is a preprocessing leak that can inflate the reported accuracy, AUC-ROC, and F1-score. The same ambiguity affects the cross-species experiment in Section 4.3: labels in the target species are relative to that species' own distribution, so the Zea-trained 92.30% is a ranking-within-species statement rather than absolute expression prediction. No table cell or repository line is given to locate the 92.30% value, and Figure 3 has no error bars. The relative ranking of models may survive this issue, but the headline 92.3% and the \"cross-species generalization\" interpretation are not established from the text alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents DeepPlantCRE, a Transformer-CNN hybrid for classifying plant gene expression from promoter, TSS-proximal, and terminator-proximal DNA sequences. The method is evaluated in within-species chromosome-level cross-validation on five plant species and eight diploid Gossypium genomes, and in cross-species transfer experiments, with reported improvements over DeepCRE and PhytoExpr and a peak cross-species accuracy of 92.3%. The authors also use DeepLIFT and TF-MoDISco to extract sequence motifs and compare them with known transcription factor binding sites in JASPAR.","tokens_in":13199,"tokens_out":7329,"duration_ms":74601,"significance":"The problem of predicting expression from cis-regulatory sequence and testing cross-species generalization is relevant, and the manuscript assembles a broad multi-species benchmark, uses external baselines and an external motif database, and makes code available. These are genuine strengths. However, as presented, the evaluation protocol contains an unresolved label-leakage risk and an ambiguity in the label definition (binary versus three-class), and the cross-species experiment lacks baseline comparisons. The quantitative claims are therefore not yet established, although they are in principle fixable with a revised evaluation.","major_comments":[{"comment":"Section 4.1 (Data Processing) says expression labels are assigned by the 25th/75th percentiles of logMaxTPM 'within each dataset,' and Section 4.2 sets k equal to the number of chromosomes. The text never states that the percentile thresholds are recomputed on the training chromosomes only. If the thresholds are estimated from the full species before the split, then the label of every gene on a held-out chromosome is determined using expression values from training chromosomes; this is a preprocessing leak that can inflate Accuracy, AUC-ROC, and F1-score. Please clarify whether thresholds are fold-specific, and if they are not, rerun the evaluation with thresholds computed on training folds only.","section":"Section 4.1, 4.2"},{"comment":"Section 4.3 reports that the Zea-trained model reaches 92.30% cross-species accuracy, but the text points to no table or figure cell with that value, and Figure 3 is a heatmap with no error bars or numeric annotations. In addition, the target-species labels in the cross-species setting are relative to that species' own logMaxTPM distribution (Section 4.1), so the 92.30% measures consistency of within-species expression ordering, not absolute expression-level prediction. The 'cross-species generalization' interpretation requires either absolute expression labels or an explicit statement that target labels are defined by target-species quantiles; please provide the exact training/test species pair and the precise protocol.","section":"Section 4.3, Figure 3"},{"comment":"Section 4.1 creates three expression classes (-1 for low, 0 for medium, 1 for high), while Section 3.3(iii) uses a sigmoid and BCEWithLogitsLoss for binary classification. The treatment of medium-expression genes is never specified: are they excluded, merged with one of the other classes, or modeled as a third class? This ambiguity affects every reported metric and the comparison with DeepCRE and PhytoExpr, so it must be resolved before the results can be interpreted.","section":"Section 3.3(iii), 4.1"},{"comment":"A stated contribution is improved cross-species generalization (Introduction and RQ2), but Section 4.3 reports transfer results only for DeepPlantCRE; no DeepCRE or PhytoExpr results are shown for the same transfer task. Without matched baselines under the same protocol, the claim that the hybrid architecture generalizes better across species than existing models is not tested.","section":"Section 4.3"},{"comment":"Equation (1) defines the input alphabet as {A,C,G,T,N} but sets X in {0,1}^{L x D} with D equal to the number of standard base types; the representation of N is unspecified. Since plant genome assemblies contain ambiguous bases in promoter and terminator regions, the actual input encoding and any masking or imputation for N must be described for the experiments to be reproducible.","section":"Equation (1), Section 4.1"}],"minor_comments":[{"comment":"The Abstract and Figure 1 list 'Arabidopsis thaliana' twice and omit 'Zea mays' from the species list, even though Zea mays is used in the experiments and is central to the reported 92.3% accuracy; Section 4.1 lists the five species correctly.","section":"Abstract, Figure 1"},{"comment":"Baseline citations are inconsistent: Section 4.1 identifies DeepCRE as [20] and PhytoExpr as [13], but Section 4.2 refers to 'DeepCRE [13]' and 'PhytoExpr [20]'; please correct the citation mapping.","section":"Section 4.1, 4.2"},{"comment":"The description of the classification head says the fully connected layer 'eventually output a continuous value representing the predicted gene expression level,' which conflicts with the binary sigmoid classification described in Section 3.3(iii); please harmonize the terminology.","section":"Section 4.1"},{"comment":"The Figure 2 caption uses 'unique hot encoding'; this should be 'one-hot encoding'.","section":"Figure 2"},{"comment":"No statistical significance tests are reported; given the overlapping standard deviations in Tables 1 and 2, the authors should report paired tests or confidence intervals for the main comparisons.","section":"Tables 1 and 2"},{"comment":"The text contains typos such as 'pepole' for 'people' and 'firt' for 'first' in the case study; the manuscript would benefit from a careful proofreading pass.","section":"Section 4.6"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you want a status check on plant gene expression classification, not if you need a definitive benchmark. The genuinely useful part is the specific stack: Transformer layer first, then residual CNN blocks, with a very small embedding, evaluated on five species plus eight Gossypium genomes. That coverage is wider than most papers in this space, and DeepCRE and PhytoExpr are the right baselines. The ablation and hyperparameter sensitivity are standard but honestly reported. The architecture is simple and plausible, and the anonymous code link is a plus.\n\nThe soft spots are real. The headline 92.3% cross-species accuracy is not verifiable from the text. The data-processing section says expression labels come from the 25th and 75th percentiles of logMaxTPM “within each dataset,” and cross-validation is then by chromosome. It never says the thresholds are recomputed on training chromosomes only. If they are not, held-out labels are partly determined by training expression values, which inflates accuracy, AUC-ROC, and F1. That is a load-bearing ambiguity, not cosmetic. The cross-species labels are also relative to each target species’ own distribution, so 92.3% is a within-species ranking statement rather than absolute cross-species expression prediction. There is no table cell or repository line pointing to that value, and Figure 3 has no error bars.\n\nSmaller problems: the input alphabet allows N but the one-hot description only covers standard bases; some passages describe three classes while the loss is binary; DeepCRE and PhytoExpr citations look shuffled; and the abstract lists Arabidopsis twice, presumably meaning Zea. These are fixable but make the paper feel rushed. The relative ordering of the three models may survive the label concern, since the same preprocessing was probably applied to all models, but the absolute numbers and the “cross-species generalization” interpretation are not established from the text alone.\n\nI would send this to peer review rather than desk reject. The modeling idea is reasonable, the dataset combination is valuable, and the issues are addressable with per-fold label construction, error bars, and a reproducible code release. I would not accept the current version, and I would not cite it in my own work until the scripts reproduce the tables and the leakage question is resolved.","headline":"A plausible architecture and a useful multi-species evaluation, but the headline 92.3% cross-species number is not established because the label construction may leak and the text never pins down where the number comes from.","tokens_in":13749,"tokens_out":3519,"would_cite":false,"duration_ms":38833,"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":"DeepPlantCRE predicts plant gene expression from DNA sequence alone and transfers across species with up to 92.3% accuracy.","keywords":["gene expression prediction","cis-regulatory elements","Transformer-CNN hybrid","cross-species generalization","plant genomics","transcription factor binding sites","DeepLIFT","TF-MoDISco"],"falsifier":"Recompute the same chromosome-held-out cross-validation on one species, but compute the low/medium/high expression cutoffs from the training folds only; if accuracy drops by more than a few points or the ranking of models changes, the reported metrics depend on test-set information. Alternatively, shuffle gene labels and confirm that DeepPlantCRE accuracy approaches chance in the cross-species setting.","tokens_in":12724,"feed_emoji":"🧬","tokens_out":5208,"duration_ms":47645,"temperature":0.7,"pith_summary":"This paper proposes DeepPlantCRE, a deep learning model that predicts whether a plant gene is expressed at low, medium, or high level using only the DNA sequence around the gene's start and end. The authors claim the model beats two existing baselines, DeepCRE and PhytoExpr, on accuracy, AUC-ROC, and F1-score across five crop species and eight cotton genomes, and that models trained on one species can predict expression in other species with accuracy up to 92.3%. The aim is to make cis-regulatory element discovery more reliable and transferable, so that regulatory variants and enhancers can be identified from sequence alone and used in crop breeding.","feed_headline":"Plant gene expression predicted from DNA by Transformer-CNN model","feed_subtitle":"Beats DeepCRE and PhytoExpr on five crops; transfers to new species with up to 92.3% accuracy.","key_machinery":"The central object is the hybrid architecture that places a single Transformer encoder layer (one attention head, embedding dimension 4) before three residual one-dimensional convolutional blocks (kernel size 8; channels 64, 128, 32), followed by a fully connected classification head with sigmoid output. The Transformer first models global sequence context over the entire promoter region; the CNN blocks then extract hierarchical local motif features, and batch normalization, dropout, learning-rate scheduling, and early stopping are used to prevent overfitting. This ordering is the paper's main design claim: global context before local feature extraction yields better cross-species generalization than either architecture alone.","core_discovery":"DeepPlantCRE, a Transformer-CNN hybrid architecture, establishes a new state of the art for sequence-only plant gene expression classification. On five species (Gossypium arboreum, Arabidopsis thaliana, Solanum lycopersicum, Sorghum bicolor, Zea mays) and eight diploid Gossypium genomes, it outperforms DeepCRE and PhytoExpr on accuracy (by 1.6–3.8%), AUC-ROC (by 1.9–3.0%) and F1-score (by 1.2–2.8%), reaching up to 92.3% accuracy in cross-species transfer when Zea mays is the training species. The model's attention mechanism captures long-range regulatory interactions, while residual CNN blocks extract local motifs; interpretability analysis with DeepLIFT and TF-MoDISco recovers known transcription factor binding motifs such as MYR2, TSO1 and TB1, supporting the claim that the model learned biologically meaningful regulatory sequence features.","pith_inferences":["If the chromosome-held-out evaluation is unbiased, the architecture's Transformer-then-CNN ordering could be a reusable pattern for other regulatory genomics tasks where long-range interactions dominate, such as enhancer–promoter interaction prediction.","The reported cross-species pattern—maize transfers best, Gossypium worst—suggests training-set size and phylogenetic distance jointly determine transferability; a systematic study across a plant phylogeny could turn this observation into a quantitative transferability predictor.","The use of two tissue-specific datasets (leaf) leaves open whether the model generalizes across tissues; a testable extension is to train on multi-tissue expression panels and see whether tissue-specific motifs emerge from the same architecture."],"forward_implications":["If the model's performance holds, regulatory element discovery in crops can be done directly from promoter sequence, without needing species-specific chromatin data.","Cross-species transfer accuracy up to 92.3% suggests a single model can annotate expression-relevant variants across related and even more distant species, prioritizing candidates for gene editing.","The recovered motifs matching known TFBSs (MYR2, TSO1, TB1, Zm00001d052229) imply that the model's internal representations align with known plant transcriptional regulation and could nominate novel CREs for validation.","The hybrid design's robustness to hyperparameter settings (CNN layers 4–8, kernel size 5–16, learning rate 1e-5 to 1e-3) means the architecture is stable enough for deployment."],"supporting_citations":[{"why":"Supplies the CNN+Transformer baseline (PhytoExpr) that DeepPlantCRE must beat.","marker":"[13]"},{"why":"Supplies the CNN baseline (DeepCRE) and the comparative dataset context for plant gene expression prediction.","marker":"[20]"},{"why":"Provides the eight diploid Gossypium genomes, annotations, and expression data used in the cross-species tests.","marker":"[32]"},{"why":"Provides the DeepLIFT attribution method used to score nucleotide contributions to expression predictions.","marker":"[25]"},{"why":"Provides the TF-MoDISco algorithm used to discover motifs from importance scores.","marker":"[26]"},{"why":"Supplies the JASPAR known transcription factor binding profiles for validating the discovered motifs.","marker":"[7]"},{"why":"Provides the Transformer attention mechanism at the core of the hybrid architecture.","marker":"[30]"}],"fun_headline_variants":["Transformer-CNN hybrid predicts plant gene expression, beats two baselines","Cross-species plant gene prediction hits 92.3% accuracy with hybrid model","DeepPlantCRE: AI model decodes plant regulatory DNA across species","Plant gene expression from DNA: hybrid model generalizes to new species"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that computing expression cutoffs within each dataset and holding out whole chromosomes for testing yields a fair estimate, with no leakage or distribution shift that would inflate the reported accuracies.","fun_headline_variants_meta":{"raw":{"variants":["Transformer-CNN hybrid predicts plant gene expression, beats two baselines","Cross-species plant gene prediction hits 92.3% accuracy with hybrid model","DeepPlantCRE: AI model decodes plant regulatory DNA across species","Plant gene expression from DNA: hybrid model generalizes to new species"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000743,"raw_usage":{"total_tokens":3378,"prompt_tokens":1070,"completion_tokens":2308,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":686,"completion_tokens_details":{"reasoning_tokens":2230}},"tokens_in":686,"tokens_out":2308,"duration_ms":14634,"temperature":1.0,"reasoning_tokens":2230,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:20:40.251337+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the same chromosome-held-out cross-validation on one species, but compute the low/medium/high expression cutoffs from the training folds only; if accuracy drops by more than a few points or the ranking of models changes, the reported metrics depend on test-set information. Alternatively, shuffle gene labels and confirm that DeepPlantCRE accuracy approaches chance in the cross-species setting.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CNN+Transformer baseline (PhytoExpr) that DeepPlantCRE must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CNN baseline (DeepCRE) and the comparative dataset context for plant gene expression prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the eight diploid Gossypium genomes, annotations, and expression data used in the cross-species tests."},{"cited_title":"Shrikumar, P","cited_arxiv_id":null,"evidence_quote":"Provides the DeepLIFT attribution method used to score nucleotide contributions to expression predictions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the JASPAR known transcription factor binding profiles for validating the discovered motifs."}],"review_version":1}