{"id":"877fcdcc-b4fe-4e45-aa0f-b84372b9d74f","arxiv_id":"2605.31562","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"TxFM applies masked autoencoding to RNA-seq counts on the new DiverseRNA-1.4M corpus and reports superior inductive transfer performance over foundation models trained on 100x larger atlases.","lead":"The paper introduces TxFM, a masked autoencoder model for learning gene expression representations from RNA-seq data on a curated 1.4M dataset. If the results hold, careful data curation and architecture choices may matter more than training on vastly larger corpora for biological applications like drug discovery.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Inductive evaluation fairness and baseline protocol consistency remain the least secured part of the outperformance claim","rationale":"The reader's weakest_assumption already isolates exactly the condition whose violation would invalidate the strongest_claim. Because the provided abstract contains no further technical detail that would falsify or strengthen that assumption, the load-bearing risk stays where the reader located it.","tokens_in":1664,"tokens_out":308,"duration_ms":16110,"concrete_test":"Extract the precise gene sets, sample IDs, and train/test splits used for each downstream task in the full manuscript; cross-check every test sample against the DiverseRNA-1.4M corpus for overlap; if any overlap exists or if baseline numbers were taken from original papers without re-evaluation on identical splits, recompute the headline performance deltas under the corrected protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that TxFM representations learned on DiverseRNA-1.4M transfer better than those from 100x-larger atlas models. This holds only if (a) all downstream tasks are strictly inductive with zero overlap between DiverseRNA-1.4M and any test sets, (b) the same task definitions, metrics, and splits are used for every baseline, and (c) no post-hoc model or hyperparameter selection occurred after seeing test results. The abstract asserts inductive evaluations and an architecture ablation but supplies no quantitative evidence that these three conditions were met when comparing against the larger FMs.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript introduces TxFM, a masked autoencoder for self-supervised learning of gene representations from RNA-seq count data. It curates the DiverseRNA-1.4M training corpus and reports that TxFM trained on this dataset produces representations that outperform transcriptomic foundation models trained on atlas-scale corpora more than 100 times larger, based on inductive representation learning evaluations; an architecture ablation is also presented to identify key design choices for transfer performance.","tokens_in":1793,"tokens_out":292,"duration_ms":11518,"significance":"If the outperformance claim holds under strictly inductive conditions with consistent evaluation protocols, the result would demonstrate that targeted data curation combined with appropriate architecture choices can yield stronger biological representations than scale alone, providing a counterpoint to the prevailing emphasis on ever-larger training corpora in transcriptomics foundation modeling.","major_comments":[{"comment":"Abstract: the central claim that TxFM on DiverseRNA-1.4M outperforms FMs trained on >100x larger corpora is load-bearing for the paper's contribution, yet the manuscript supplies no quantitative verification that downstream task splits have zero overlap with DiverseRNA-1.4M, that identical task definitions/metrics/splits were applied to all baselines, and that no post-hoc model or hyperparameter selection occurred after inspecting test results.","section":"Abstract"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their thorough review and for highlighting the importance of rigorous verification for our central inductive transfer claims. We address the major comment below and commit to revisions that will strengthen the manuscript's transparency and reproducibility.","responses":[{"response":"We agree that explicit verification of these points is essential to substantiate the load-bearing claim of outperformance under strictly inductive conditions. The current manuscript does not include quantitative overlap checks or a consolidated description of protocol consistency across models. In the revised version, we will add a new appendix (or expanded methods section) that: (1) provides quantitative verification of zero overlap between downstream task splits and DiverseRNA-1.4M via explicit checks on sample and gene identifiers; (2) documents the precise task definitions, metrics, and data splits used for every baseline, confirming they were applied identically; and (3) details the hyperparameter selection workflow, which relied exclusively on validation performance with no access to test results at any stage. These additions will directly address the referee's concerns while preserving the paper's focus on data curation and architecture choices.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central claim that TxFM on DiverseRNA-1.4M outperforms FMs trained on >100x larger corpora is load-bearing for the paper's contribution, yet the manuscript supplies no quantitative verification that downstream task splits have zero overlap with DiverseRNA-1.4M, that identical task definitions/metrics/splits were applied to all baselines, and that no post-hoc model or hyperparameter selection occurred after inspecting test results."}],"tokens_in":1272,"tokens_out":344,"duration_ms":16575,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main point is that TxFM, trained with masked autoencoding on their new 1.4M curated RNA-seq corpus, reportedly gives better gene representations than foundation models trained on 100x more data, at least for the inductive tasks shown.\n\nThey do a few things right. The architecture ablation identifies specific choices that help transfer, which is more useful than just reporting a new model. Releasing DiverseRNA-1.4M as a public resource lets others test whether curation really beats scale. The focus on inductive evaluation rather than just in-distribution performance also matches a real need in transcriptomics, where batch effects and noise are common.\n\nThe soft spot is the evaluation protocol. The outperformance claim only holds if the downstream tasks have zero overlap with DiverseRNA-1.4M, if every baseline used the exact same splits and metrics, and if no post-selection happened after seeing test results. The abstract does not supply those checks, so the central result stays provisional until the methods are examined.\n\nThis paper is for people building or applying transcriptomic models in drug discovery or cellular state modeling. Readers who want concrete ideas on data curation and architecture tweaks for self-supervised bio models will find it worth reading.\n\nIt deserves peer review because the question about curation versus scale is timely and the work is grounded enough to be worth referee time, even if the current evidence needs tightening.","headline":"TxFM on DiverseRNA-1.4M beats much larger atlas models on inductive tasks, but the abstract leaves the fairness of those comparisons unverified.","tokens_in":2292,"tokens_out":359,"would_cite":false,"duration_ms":14617,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A masked autoencoder on a curated 1.4 million RNA-seq samples produces higher-fidelity gene representations than foundation models trained on over 100 times more data.","keywords":["transcriptomics","self-supervised learning","masked autoencoder","RNA-seq","gene expression","representation learning","foundation models","inductive transfer"],"falsifier":"A downstream gene-expression task in which TxFM embeddings produce lower performance than a linear baseline or a model trained on a larger atlas after controlling for evaluation protocol and data overlap.","tokens_in":2591,"feed_emoji":"","tokens_out":645,"duration_ms":13866,"temperature":0.7,"pith_summary":"The paper tests whether deep representation learning adds value over raw transcript counts for RNA-seq data by building TxFM, a masked autoencoder trained with self-supervision. It curates a public dataset called DiverseRNA-1.4M and shows through inductive transfer evaluations that the resulting gene embeddings outperform transcriptomic foundation models trained on atlas-scale collections more than 100 times larger. Ablations identify specific architecture choices that enable strong downstream performance. The work concludes that inductive self-supervised learning can succeed in transcriptomics when model design and data curation are aligned carefully. This matters because existing models often fail to beat simple linear baselines despite their scale.","feed_headline":"Curated 1.4M RNA-seq samples beat 100x larger atlas models","feed_subtitle":"Masked pretraining on carefully assembled data yields stronger gene representations than scale alone.","key_machinery":"Masked autoencoding tailored to RNA-seq count data, with architecture ablations that identify configurations required for strong transfer.","core_discovery":"TxFM employs a masked autoencoding approach tailored to diverse RNA-seq count data; when trained on the curated DiverseRNA-1.4M corpus, it yields high-fidelity gene representations that outperform foundation models trained on atlas-scale corpora over 100x larger, indicating that inductive self-supervised learning is viable for transcriptomics representation provided careful synthesis of architecture and training data curation.","pith_inferences":["Drug-discovery pipelines that rely on gene-expression signatures could substitute TxFM embeddings for raw counts or larger-model outputs without increasing compute.","The same masking strategy might extend to other count-based biological modalities if the count-distribution handling is preserved.","Future work could test whether adding explicit batch-effect correction inside the masking objective further improves out-of-distribution transfer."],"forward_implications":["Inductive self-supervised learning becomes a practical route for transcriptomic representation learning.","Data curation can matter more than raw corpus scale for building effective gene embeddings.","Architecture choices identified in the ablations become necessary components for strong transfer in similar models.","Raw transcript counts are no longer the default baseline once masked pretraining is applied correctly."],"fun_headline_variants":["Curated 1.4M RNA-seq outperforms 100x larger models","Masked pretraining on 1.4M data yields superior gene representations","Smaller curated corpus outperforms atlas models 100x its size","TxFM masked model outperforms 100x larger atlases with curated data"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The inductive evaluations compare models fairly without data leakage or post-hoc selection, and the curated DiverseRNA-1.4M dataset represents the distribution of target downstream tasks.","fun_headline_variants_meta":{"raw":{"variants":["Curated 1.4M RNA-seq outperforms 100x larger models","Masked pretraining on 1.4M data yields superior gene representations","Smaller curated corpus outperforms atlas models 100x its size","TxFM masked model outperforms 100x larger atlases with curated data"]},"model":"grok-4.3","cost_usd":0.00855,"raw_usage":{"total_tokens":3846,"prompt_tokens":637,"num_sources_used":0,"completion_tokens":76,"cost_in_usd_ticks":85499500,"prompt_tokens_details":{"text_tokens":637,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3133,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":637,"tokens_out":76,"duration_ms":18418,"temperature":1.0,"reasoning_tokens":3133,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-28T23:10:31.318483+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A downstream gene-expression task in which TxFM embeddings produce lower performance than a linear baseline or a model trained on a larger atlas after controlling for evaluation protocol and data overlap.","supporting_citations":[],"review_version":1}