{"id":"0beb177d-388e-4bd3-a257-54c248c1aeb9","arxiv_id":"1908.07520","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Using genetic instruments and Hi-C contacts, the authors built causal gene networks and identified nine genes with differential regulation in schizophrenia plus candidate new risk genes.","lead":"This paper combines DNA variants, brain gene activity, and 3D chromatin contacts to build causal gene regulation maps, then compares schizophrenia patients with healthy controls. It reports nine genes with altered regulation in schizophrenia and several new candidate risk genes that could be studied as drug targets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All directed edges rest on untested Mendelian randomization instrument validity; with no F-statistics, pleiotropy screens, or adult-brain Hi-C, the 9 differential genes and 4 hubs are unidentified if instruments fail.","rationale":"The reader's REJECT verdict is appropriate. The weakest assumption I find is the instrument-validity premise underlying every directed edge. This is not a disagreement with consensus; it is an internal identification problem: the paper's own Step 1 treats genetic variation as exogenous, but no evidence is supplied that the selected variants satisfy relevance and exclusion restrictions in adult DLPFC. The fetal Hi-C data makes tissue mismatch a concrete threat. The predictive accuracy of hubs (e.g., Figure 5B) is real evidence of association, not of causation; without valid instruments, the directed interpretation remains unsupported. I also note the absence of a detailed methods section and the lack of multiple-testing correction, but these reinforce rather than replace the MR concern. If the instrument check I propose passed, the central claim would be much stronger; given the complete absence of such diagnostics, REJECT stands. No change to the reader's verdict.","tokens_in":11431,"tokens_out":3711,"duration_ms":41598,"concrete_test":"Using the CMC DLPFC genotype-expression data themselves, compute for each of the 81 reported instrument-gene pairs: the conditional F-statistic for the variant(s) on the target gene, an MR-Egger intercept test for directional pleiotropy, and a colocalization posterior (PP.H4) against adult-DLPFC eQTL evidence. Re-run the G-DAG and the case/control comparison retaining only variants with F>10, MR-Egger intercept p>0.05, and PP.H4>0.8, then check whether the 9 differentially regulated genes and the TENM3/NRXN3/MYH10/PEX5L hub modules survive. If the edge set or the gene lists change materially, the causal claims are not identified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the network edges are causal and that GABRA2, LRRTM2, PPM1E, etc. are differentially regulated in SCZ—depends entirely on Step 1's assertion that 'genetic inherited variation is the cause of gene transcription not the other way around.' For this to identify directions, each selected variant must be a valid instrument for its gene's transcription: relevant (strongly associated with the target gene), independent of confounders, and affecting other genes only through the target (no horizontal pleiotropy). None of these are checked. Only '81 showed significant effect' is reported, with no F-statistics, no p-value threshold, no colocalization with eQTLs, and no pleiotropy screen. The Hi-C contacts used to select variants are from mid-gestation fetal cortex, not adult DLPFC where the RNA-seq was measured; regulatory architecture differs across development. A variant that directly affects a downstream gene or is in LD with another causal variant would create a spurious directed edge from the target to that gene. Every downstream conclusion—the 9 differential genes, the four essential hubs, and the novel SCZ candidates—is built on those edges. The manuscript also defers to a 'method section' for details, but no such section with MR diagnostics is present in the text.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an approach that integrates RNA-seq, genotype, and Hi-C data to build \"transcriptomic causal networks\" via a Mendelian randomization and graphical-modeling framework adapted from the G-DAG algorithm. Applied to CommonMind Consortium dorsolateral prefrontal cortex data, the approach is used to identify differential gene regulatory patterns between schizophrenia cases and controls, to define modules around hubs (TENM3, NRXN3, MYH10, PEX5L) claimed to be essential for brain function, and to propose five novel schizophrenia candidate genes (RNF150, UNC5D, RTF1, LMO7, SEZ6L). The abstract reports nine genes differentially regulated in schizophrenia, including GABRA2, LRRTM2, and PPM1E. The paper is written as a methods-oriented study, but no methods section is included in the provided text.","tokens_in":11687,"tokens_out":5017,"duration_ms":50179,"significance":"If the causal-network estimates and the differential-regulation claims were valid, this would be a substantive contribution to the mechanistic understanding of schizophrenia and to the identification of potential therapeutic targets. The integration of germline genetic variation, gene expression, and chromatin conformation into directed networks is an important direction, and the use of a public, well-characterized dataset is a strength. However, as presented, the central claims are not statistically established: the causal directions rest on untested instrumental-variable assumptions, the differential-regulation results lack formal tests and multiple-testing control, and the module and novel-gene discoveries are conditional on the known schizophrenia-associated gene set used to construct the network. The manuscript also lacks the methodological detail needed for reproducibility.","major_comments":[{"comment":"The causal interpretation of every directed edge rests on the validity of the selected genetic variants as instruments for gene transcription. The manuscript reports only that \"81 showed significant effect on gene transcription level\" among 4,641 variants, with no F-statistics, no minimum effect-size or p-value threshold, no colocalization with eQTL evidence, and no horizontal-pleiotropy screen. Moreover, the Hi-C contacts used to select variants come from mid-gestation fetal cortex (Won et al., 2016), whereas the RNA-seq data are from adult DLPFC, and regulatory architecture differs across development; a variant that affects multiple genes directly, or is in linkage disequilibrium with another causal variant, would produce spurious directed edges. Because all downstream conclusions, including the nine differentially regulated genes and the four hubs, are built on these edges, the absence of instrument diagnostics is a load-bearing gap.","section":"Step 1 (Building transcriptomic causal network) and Results"},{"comment":"The central claim that GABRA2, LRRTM2, PPM1E, SORT1, GNAL, ZNF692, ZNF672, SNRNP48, and RALGPS2 are \"differentially regulated\" in schizophrenia is read from forest plots showing mean effect sizes with 95% confidence intervals for cases versus controls. No formal statistical test of the case-control difference is reported, no p-values accompany the comparisons, and no multiple-testing correction is applied across the nine genes or across the many edges in the network. Overlapping confidence intervals do not establish equivalence, and non-overlap is not a test of significance, so the differential-regulation claim is not statistically supported.","section":"Results, \"Differential gene regulatory patterns\" and Figures 3-4"},{"comment":"The module identification and novel-gene discovery are partly circular. The 1,181-gene network is explicitly constructed from \"a class of genes that includes the highest number of SCZ-associated genes,\" and the novel candidates (RNF150, UNC5D, RTF1, LMO7, SEZ6L) are inferred under the assumption \"that the SCZ-associated genes are truly related to SCZ.\" Consequently, the finding that modules are enriched for schizophrenia-associated genes and that novel candidates are connected to known schizophrenia genes is shaped by the prior labels rather than being an independent discovery. No independent replication or external functional validation is provided.","section":"Step 4 and Results, \"Modules\" and novel-gene predictions"},{"comment":"The text repeatedly defers to a \"method section\" for the G-DAG adaptation, edge p-value cutoffs, degree thresholds, penalization parameters, cross-validation schemes, and conditional-analysis details (Steps 1, 3, and 4; TENM3 module). No such section is present in the manuscript. Without these details, in particular the instrument selection protocol and the structure-learning algorithm, the results cannot be reproduced or independently assessed. This is not purely a presentation issue; it blocks verification of the central causal claims.","section":"Methods (missing)"},{"comment":"The classification of genes as broadcasters, receptors, or mediators depends on thresholds for out-degree and in-degree that are described only as \"a given out-degree\" and \"a given in-degree.\" These thresholds are never specified, and no sensitivity analysis is reported for their choice. Because this classification drives the identification of hubs, mediators, and differential regulatory patterns, the missing criteria and sensitivity analysis constitute a substantive gap.","section":"Step 2 (Identification of receptor, broadcaster, and mediator genes)"}],"minor_comments":[{"comment":"The abstract lists \"ZNF69\" but the Results section names \"ZNF692\"; these should be reconciled (the gene symbol appears to be ZNF692, as used in the body text).","section":"Abstract and Results"},{"comment":"\"RNAF150\" appears in the abstract but the gene is referred to as RNF150 in the Results; please correct the typo.","section":"Abstract"},{"comment":"The module numbering is inconsistent: two rows are both labeled \"3\" (MYH10 and PEX5L); renumber the modules uniquely.","section":"Table 1"},{"comment":"The text contains an unresolved reference \"[Ref]\" when discussing the NRXN3 protein family; this citation needs to be completed.","section":"NRXN3-Module"},{"comment":"The text uses the typo \"MHYH10\" in the discussion of the MYH10 prediction results; this should be corrected to MYH10.","section":"MYH10-Module"},{"comment":"The manuscript references Supplementary Tables 1-13 and Supplementary Figures 1-2, but no supplementary material is provided with the submitted text; please ensure all supplements are included.","section":"Supplementary materials"}],"recommendation":"reject","confidential_remarks":"This manuscript is at an early stage and does not currently meet the standards for publication. The most serious issues are the absence of any methods section and the lack of instrumental-variable diagnostics for the Mendelian randomization step, which undermines the identification of all directed edges. The differential-regulation claims are also unsupported by formal statistical tests. I would not rule out a future resubmission if the authors add a complete methods section, instrument strength and pleiotropy checks (ideally with adult brain Hi-C or other adult regulatory annotations), formal case-control tests with multiple-testing correction, and a de-biased or externally validated approach to novel gene discovery. As it stands, the central causal claims are not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat you should know about this one: it's the authors' G-DAG method, previously used on metabolomics, applied to schizophrenia postmortem brain. The new bits are the Hi-C-informed variant selection and a case/control comparison of network modules. Nine mediator genes and four hubs are the headline results. The method is interesting, but the paper doesn't support its own strong conclusions.\n\nWhat it does well: the pipeline integrates genotype, RNA-seq, and Hi-C data into directed networks, which is a step beyond correlation-based gene co-expression. The authors check predictions with cross-validation, and they use external databases (HPA, STRING) to give biological context. That's more than most network papers do.\n\nThe soft spots are serious. The causal directions are the foundation of everything, and they rest on Mendelian randomization assumptions that are never tested. We don't get F-statistics or any weak-instrument check for the 81 variants; no pleiotropy screen; and the Hi-C data are from fetal cortex, not adult DLPFC, so the regulatory architecture may not match the expression data. If any variant affects a downstream gene directly or via linkage, the directed edges and everything built on them are spurious.\n\nThe differential regulation claim is also underpowered. The \"9 genes differentially regulated\" come from forest plots with overlapping confidence intervals, with no formal test or multiple-testing correction. That's a central result, and it needs a proper interaction test.\n\nThere's also a circular flavor: the network is built on 1,181 transcripts chosen for enrichment of known SCZ-associated genes, and the novel candidates are then predicted from their position relative to those same genes. Conditional analysis is a reasonable way to generate hypotheses, but it's not independent evidence.\n\nFinally, the manuscript refers to a method section that isn't present, and there's no code. That makes it hard to evaluate or reproduce.\n\nSo my take: as a hypothesis-generating systems biology paper, it's worth reading. As a claim to have identified causal regulators in schizophrenia, it's not convincing in this form. I'd send it to peer review because the approach and the candidate gene list could be useful to the field, but the authors need to add MR diagnostics, a formal test for differential regulation, and better validation of the novel genes before I'd believe the results.","headline":"A promising adaptation of G-DAG to brain transcriptomics, but the causal and differential claims outrun the evidence.","tokens_in":12236,"tokens_out":2612,"would_cite":false,"duration_ms":26918,"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":"Transcriptomic causal networks built from genetic, RNA, and 3D chromatin data identify nine genes differentially regulated in schizophrenia and four hubs essential for normal brain function.","keywords":["schizophrenia","transcriptomic causal network","gene regulation","Mendelian randomization","Hi-C","RNA-seq","differential gene regulation","essential genes"],"falsifier":"A pleiotropy check on the 81 variants with significant transcription effects would settle the edge directions: if a variant assigned to one gene also associates with downstream genes after conditioning on that gene, or if instrument strength is weak (F-statistic below 10), the causal orientation is not identified. A second decisive check is whether adult prefrontal Hi-C contacts reproduce the fetal contacts used to select the instruments.","tokens_in":11238,"feed_emoji":"🧠","tokens_out":7485,"duration_ms":68503,"temperature":0.7,"pith_summary":"Schizophrenia is usually studied one gene at a time, but this paper treats the brain's transcription program as a causal network and asks which regulatory connections change between cases and controls. By combining genetic variants, RNA-seq from post-mortem prefrontal cortex, and 3D chromatin contact data, the authors build directed edges from the principle that genetic variation causes transcription, and identify nine genes whose regulation differs in schizophrenia: GABRA2, LRRTM2, PPM1E, SORT1, GNAL, ZNF69, RALGPS2, ZNF672, and SNRNP48. They also find four hubs, NRXN3, TENM3, MYH10, and PEX5L, that predict downstream transcription in both cases and controls, calling them essential for brain function, and propose five new candidate schizophrenia genes. A sympathetic reader would care because this moves from a list of differentially expressed genes to a causal wiring diagram with specific intervention points.","feed_headline":"Nine genes show altered regulation in schizophrenia brains","feed_subtitle":"A causal network from genetics, RNA, and 3D chromatin also marks four hubs essential for normal brain function.","key_machinery":"The machinery is the transcriptomic causal network, a directed acyclic graph in which an edge from one gene to another means a change in the first gene's transcription causes a change in the second's. Edges are oriented by Mendelian randomization: for each gene, genetic variants are selected from exome-chip SNPs and fetal-brain Hi-C contacts, and because alleles are assigned randomly at conception, variation in the target gene's transcription is treated as a cause, not an effect, of downstream genes. The paper adapts the G-DAG algorithm to build this graph, then classifies genes by in- and out-degree into broadcasters, receptors, and mediators, defines modules around high-out-degree hubs, and tests hub-to-downstream predictions with penalized regression and cross-validation.","core_discovery":"The central claim is that a transcriptomic causal network, built by treating inherited genetic variation as the cause of transcription and using Hi-C contacts to pick cis-acting variants, can separate regulatory wiring that is preserved in schizophrenia from wiring that breaks. Applied to 1,181 transcripts from post-mortem dorsolateral prefrontal cortex, the network yields nine mediator genes whose effect on downstream genes is lost in SCZ cases: GABRA2, LRRTM2, PPM1E, SORT1, GNAL, ZNF69, RALGPS2, ZNF672, and SNRNP48. Three of these, GABRA2, LRRTM2, and PPM1E, are preferentially expressed in brain and related to synaptic function; GNAL sits between two antipsychotic drug targets. Four hubs, NRXN3, TENM3, MYH10, and PEX5L, predict the transcription of their downstream genes in both cases and controls, so the paper labels them essential for brain function; they anchor modules containing many schizophrenia-associated genes. The same logic predicts RNF150, UNC5D, RTF1, LMO7, and SEZ6L as new schizophrenia-associated genes. The paper also observes that highly connected genes are less likely to be influenced directly by genetic variants, implying that schizophrenia risk acts largely through trans-regulation rather than cis effects on hub genes.","pith_inferences":["Editorial inference: if the instrument-validity assumption fails, the whole edge set is unidentified; a natural next paper would re-build the network with adult prefrontal Hi-C data and report instrument strength statistics.","Editorial inference: the hubs-and-modules structure makes a crisp perturbation prediction: CRISPR interference against NRXN3, TENM3, MYH10, or PEX5L in human neurons should move downstream genes in the signed directions shown, testable without relying on Mendelian randomization assumptions.","Editorial inference: the observation that high-connectivity genes are less cis-regulated may generalize beyond schizophrenia; applying the same pipeline to another brain disorder or tissue would show whether it is a general feature of gene regulatory networks."],"forward_implications":["The nine differentially regulated genes are candidate causal mediators of schizophrenia-related transcription, not just markers, and are testable in independent post-mortem or stem-cell datasets.","NRXN3, TENM3, MYH10, and PEX5L should be treated as hubs whose disruption would alter many downstream genes, so therapies aimed at downstream targets may avoid breaking essential brain regulation.","RNF150, UNC5D, RTF1, LMO7, and SEZ6L are predicted schizophrenia genes that can be prioritized for targeted sequencing or functional assays.","Because highly connected genes are less likely to carry direct cis-regulatory variant effects, genetic risk for schizophrenia is expected to concentrate in trans-regulation of peripheral genes rather than in hubs themselves."],"supporting_citations":[{"why":"Supplies the G-DAG algorithm that the transcriptomic causal network adapts from metabolomics.","marker":"Yazdani et al., 2016"},{"why":"Provides the RNA-seq data and normalization from post-mortem dorsolateral prefrontal cortex and the schizophrenia-associated gene list used for mapping.","marker":"Fromer et al., 2016"},{"why":"Supplies the fetal-brain Hi-C chromatin contacts used to select cis-regulatory genetic variants.","marker":"Won et al., 2016"},{"why":"Supplies the Hamming-distance assessment and network-structure learning machinery used to evaluate the fitted graph.","marker":"Tsamardinos, Brown, & Aliferis, 2006"},{"why":"Provides protein-protein interaction evidence used to support the regulatory relationships and module assignments.","marker":"Szklarczyk et al., 2019"},{"why":"Provides tissue-specific mRNA and protein expression evidence used to argue brain relevance of the identified genes.","marker":"Uhlén et al., 2015"}],"fun_headline_variants":["Causal network finds nine genes with altered regulation in schizophrenia brains","Network reveals nine gene regulation breaks in schizophrenia","Causal network identifies nine genes with lost regulation in SCZ","Nine genes lose regulation in schizophrenia, causal network shows","Causal network predicts five new genes for schizophrenia"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every directed edge assumes the genetic variants assigned to a gene affect that gene's transcription and reach other genes only through it, with no sideways effects and no hidden confounders, and the 3D chromatin contacts used to pick those variants come from fetal brain, not the adult prefrontal cortex being studied.","fun_headline_variants_meta":{"raw":{"variants":["Causal network finds nine genes with altered regulation in schizophrenia brains","Network reveals nine gene regulation breaks in schizophrenia","Causal network identifies nine genes with lost regulation in SCZ","Nine genes lose regulation in schizophrenia, causal network shows","Causal network predicts five new genes for schizophrenia"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000781,"raw_usage":{"total_tokens":3485,"prompt_tokens":1016,"completion_tokens":2469,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":2392}},"tokens_in":632,"tokens_out":2469,"duration_ms":16674,"temperature":1.0,"reasoning_tokens":2392,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:18:23.915480+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A pleiotropy check on the 81 variants with significant transcription effects would settle the edge directions: if a variant assigned to one gene also associates with downstream genes after conditioning on that gene, or if instrument strength is weak (F-statistic below 10), the causal orientation is not identified. A second decisive check is whether adult prefrontal Hi-C contacts reproduce the fetal contacts used to select the instruments.","supporting_citations":[],"review_version":1}