{"id":"b148993e-dd10-43dc-bf34-dcf4a92e0c5e","arxiv_id":"2508.07427","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An expanded RNA-centered knowledge graph integrating around 100 million interaction records from 91 sources, with nodes and edges enriched by biological context, sequences, and standardized identifiers.","lead":"RNA-KG v2.0 is a new version of a knowledge graph that maps how RNA molecules interact with genes, proteins, diseases, and chemicals, now with about 100 million relationships and contextual details such as tissue or cell line. It aims to make it easier for researchers to query RNA biology and to train AI models that predict new interactions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No error-rate evaluation for the look-up tables that map source identifiers to RNAcentral/Ensembl/OBO terms; if a meaningful fraction is wrong, conflated nodes/edges infect every query and link-prediction result.","rationale":"The reader's weakest assumption correctly identifies the identifier look-up tables as the foundation of the entire resource. The paper gives no precision/recall evaluation for v2.0 mappings, even though 20 new sources were added and the RNAcentral/Ensembl scheme is new. If those mappings are wrong, then the node counts, edge counts, properties, and all downstream queries/predictions are built on unreliable entity resolution. This is more load-bearing than the link-prediction leakage concern, because the leakage affects only one application result, whereas mapping errors corrupt the resource itself and every use case. The concern is also concretely testable because the Neo4j endpoint, API, and code are public. I therefore agree with the reader's conditional verdict: the resource is plausible and reproducible, but the missing mapping validation plus the overstatement of 'manually curated' for prediction-derived edges should be corrected or clarified before full acceptance.","tokens_in":17470,"tokens_out":4349,"duration_ms":50107,"concrete_test":"Using the public Neo4j endpoint/API, export all relationships contributed by one new prediction-heavy source (e.g., miRanda or circBase) together with their original source IDs and mapped RNAcentral/Ensembl IDs. Draw a stratified random sample of at least 500 mapped pairs and independently re-map each original ID via the source's current API and RNAcentral/Ensembl ID-mapping services, comparing sequences/genomic coordinates. Also query a sample of source IDs absent from the look-up table to estimate recall. Report exact-match precision with a 95% CI; if precision is not ≥99% or recall is not reported, the 'standard identification schemes' claim is unsupported and downstream link-prediction/numerical claims need re-auditing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that RNA-KG v2.0 is a reliable property graph with 6,553,767 nodes and 99,936,712 relationships grounded in standard identification schemes. The load-bearing step is the identifier mapping: 'Look-up tables were used to map proprietary identifiers to standard identification schemes as shown in [10]' (New Data Sources and Identification Schemes). No precision, recall, or manual audit of these tables is reported, despite the paper stating that 'manual efforts are often required to retrieve standard identifiers.' Because non-ontological entities are linked to ontology classes through the 'entity linking technique according to the class-based modeling described in [7]', a wrong mapping or wrong subclass assignment does not just add a bad edge; it conflates distinct molecules (e.g., two isoforms collapsed, or a proprietary ID matched to the wrong RNAcentral URS) and propagates through rdfs:subClassOf hierarchy into every Cypher query, view, and link-prediction experiment. The problem is compounded by labeling all ~100M relationships 'manually curated' when several new sources (miRanda, RNAhybrid, microT, STRING) are computational predictions; even if each source is reputable, those edges are not the same evidentiary class as curated databases, and their identifiers may be less stable. The resource is public, so the gap is testable, but as written the paper provides no evidence that the mapping error rate is low.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces RNA-KG v2.0, a property-graph knowledge base centered on RNA interactions. It integrates 80 repositories and 11 ontologies, yielding 6,553,767 nodes and 99,936,712 relationships, with 27,242,075 node properties and 242,896,588 edge properties. The authors describe a new identification scheme based on RNAcentral and Ensembl, the addition of 20 new data sources, the enrichment of nodes and edges with context properties (tissue, cell line, method, PubMed IDs), and a public Neo4j endpoint, web portal, and REST API. Several use cases are presented, including context-aware Cypher queries, content-aware pruning, and topological and multimodal link prediction on a miRNA-disease view, plus a time-stratified link-prediction validation. The core contribution is a large, publicly accessible RNA-centered KG with standardized identifiers and rich properties.","tokens_in":17764,"tokens_out":3259,"duration_ms":38761,"significance":"If the data-quality concerns below are addressed, RNA-KG v2.0 would be a valuable resource for the RNA bioinformatics community. The scale (100M relationships, hundreds of edge types, millions of properties) is substantially larger than the previous version and enables queries that most other RNA interaction resources do not support, especially context-aware filtering by method, tissue, disease, and supporting literature. The grounding of non-coding RNAs in RNAcentral and mRNAs in Ensembl is a welcome improvement over gene-level conflation, and the public portal/API and reproducible data release (Zenodo/GitHub) are concrete strengths that allow independent validation. However, the claims as written are weakened by a misleading characterization of all interactions as 'manually curated' and by the absence of an evaluation of the identifier-mapping step, which is load-bearing for any KG claiming standard identification schemes.","major_comments":[{"comment":"The abstract and introduction describe the ~100M interactions as 'manually curated'. However, the newly integrated sources listed in the same section include miRanda, RNAhybrid, microT (computational target predictors), STRING (computational association predictions), and GeneMANIA (functional association predictions). Those edges are not manually curated experimental observations and have different reliability profiles. Conflating them inflates the apparent evidence base and is not a purely terminological issue: downstream users will interpret 'manually curated' as experimental validation. Please reclassify interactions by evidence type (curated, experimental, computational prediction), report counts per category, and adjust the abstract accordingly. The edge property 'score' mentioned in 'Edges’ properties' is not systematically described; clarify how prediction scores from sources such","section":"Abstract and New Data Sources and Identification Schemes"},{"comment":"The statement 'Look-up tables were used to map proprietary identifiers to standard identification schemes as shown in [10]' is the sole support for the identifier-mapping step. The paper reports no precision, recall, or manual audit of these tables. Because a wrong mapping can conflate distinct isoforms (e.g., two RNAcentral URS entries) or attach an edge to the wrong entity, a non-negligible error rate would propagate through every Cypher query, view, and link-prediction experiment. The paper itself acknowledges that 'manual efforts are often required to retrieve standard identifiers,' yet provides no evidence that the large-scale mapping is accurate. This is testable: the public endpoint allows sampling and verifying mappings against RNAcentral/Ensembl cross-references. Please add a mapping-quality evaluation (or at minimum a clear limitation section and an error estimate from a manual","section":"New Data Sources and Identification Schemes (Look-up tables)"},{"comment":"The pruning experiment reports 'reduced to 9.3k nodes and 196.3M edges' from an initial subgraph of approximately 9.8k nodes and 244.8k edges. The figure '196.3M' cannot be correct; a 19.8% reduction from 244.8k yields approximately 196.3k edges. This appears to be a typo, but as written it describes a graph four orders of magnitude larger than the input, which would undermine the quantitative claim if taken literally. Please correct and verify all numbers in this example.","section":"Applications and Use Cases, Example 5 (Fig. 9)"}],"minor_comments":[{"comment":"The abstract says '91 linked open data repositories and ontologies' while the introduction says '80 publicly available repositories and 11 biomedical ontologies.' This is consistent arithmetically, but the phrasing may confuse; consider writing '80 repositories and 11 ontologies (91 sources)' for clarity.","section":"Abstract"},{"comment":"Typo: 'synomyms' should be 'synonyms'.","section":"RNA-KG Bio-Entities (Bio-entities’ properties)"},{"comment":"The phrase 'according to the proposed histograms' is unclear; presumably 'according to the identification schemes listed in Table 1' is intended.","section":"RNA-KG Bio-Entities (node type distribution)"},{"comment":"The statement that 82% of predicted links achieve scores above 0.5 is reported without a random or negative-control baseline. As a validation of KG quality, it would be more convincing if compared to a null distribution (e.g., scores on a permuted graph). At the very least, note that high raw scores may reflect model calibration rather than predictive signal.","section":"Applications and Use Cases (Time-stratified link prediction, Fig. 11)"},{"comment":"The choice to represent ternary relationships as lists of attributes is justified, but the paper should explicitly warn that querying with IN or list containment may be less efficient and that users seeking reified n-ary structures must transform the data themselves. This is a usability caveat worth stating.","section":"Edges’ properties"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be of interest to the database and bioinformatics communities, but the 'manually curated' claim in the abstract is a potential overstatement that reviewers and readers will notice. The absence of a mapping-error evaluation is a substantive gap for a resource whose central purpose is reliable integration. I do not see evidence of deliberate misrepresentation; the issues appear to be omissions and imprecise wording, and the public availability of the data makes them correctable in a revision. If the authors can provide evidence-type breakdowns, a mapping audit, and corrections to the pruning numbers, the paper could be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RNA-KG v2.0 is a real resource release: a property-graph version of the RNA knowledge graph with ~6.5M nodes and ~100M edges, twenty new sources, isoform-level identifiers via RNAcentral/Ensembl, and node/edge properties that enable context-aware queries. The public Neo4j endpoint, API, code, and data on Zenodo/GitHub make it testable, and the Cypher examples are concrete. That part deserves credit.\n\nThe soft spots are real but not fatal. The abstract's \"manually curated\" label is wrong for a meaningful share of the edges. New sources include miRanda, RNAhybrid, microT, STRING, and GeneMANIA, which are computational prediction databases. Those edges are not the same evidentiary class as curated ones, and the paper doesn't distinguish them except by source tag. More importantly, the identifier mapping is load-bearing: proprietary IDs are mapped via look-up tables with no precision, recall, or manual audit reported. A bad mapping conflates isoforms or attaches an interaction to the wrong molecule, and the error propagates through the ontology hierarchy. The resource being public means the gap is fixable, but as written it's a hole.\n\nThe link prediction use cases are the weakest section. The time-stratified setup trains on pre-2022 edges but appears to use node2vec embeddings learned on the full graph, which would leak post-2022 information; at minimum the paper doesn't say the embeddings were trained only on pre-2022 edges. Table 2 has no error bars or significance tests, and the multimodal gains are mixed. Again, these flaws don't sink the resource—the KG is the contribution, not the prediction benchmarks.\n\nWho gets value: anyone building or using RNA-centric knowledge graphs, or doing KG integration with biomedical ontologies. It's a solid dataset paper, not a methods breakthrough. I'd send it to peer review with a request to fix the provenance language, report identifier-mapping accuracy, and tighten or flag the link prediction section as preliminary.","headline":"A useful, testable resource release with a few provenance and evaluation gaps that need fixing before I'd trust the numbers.","tokens_in":18264,"tokens_out":2062,"would_cite":true,"duration_ms":22666,"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":"RNA-KG v2.0 is a property-graph knowledge base that integrates about 100 million curated RNA-centered interactions from 91 linked open-data sources and attaches context properties—cell line, tissue, disease, experimental method, and support","keywords":["knowledge graph","RNA interactions","property graph","biomedical ontologies","data integration","non-coding RNA","link prediction","context-aware querying"],"falsifier":"Independently re-resolve a random sample of 1,000 RNA-KG v2.0 edges: take both endpoint identifiers back to the original source records and to the standard scheme (RNAcentral, Ensembl, or ontology term) via the published look-up tables, and count mismatches. A mismatch rate materially above zero, or a failure to reproduce the claimed 6,553,767 nodes and 99,936,712 relationships from the released code and input files, would falsify the construction claim.","tokens_in":17348,"feed_emoji":"🧬","tokens_out":8379,"duration_ms":85260,"temperature":0.7,"pith_summary":"The paper sets out to upgrade RNA-KG from a plain ontology-based graph of RNA-centered associations to a property graph in which both molecules and their interactions carry structured attributes. The central claim is that this added context—sequence, genomic coordinates, cell line, tissue, pathological state, experimental method, and supporting literature—makes the roughly 100 million integrated interactions substantially more usable for fine-grained queries and for machine-learning tasks such as context-aware link prediction. A sympathetic reader would take the contribution to be a resource and a demonstration: the paper shows concrete Cypher queries, a content-aware pruning experiment, and link-prediction results that improve when node properties are added. If the resource is reliable, it lowers the barrier to asking condition-specific questions about RNA biology.","feed_headline":"100M RNA interactions now come with context and evidence","feed_subtitle":"Property-graph release lets researchers filter 91 sources by tissue, cell line, disease, and method.","key_machinery":"The property graph model is the load-bearing mechanism: nodes carry labels, descriptions, synonyms, sequences, genomic coordinates, and species, while edges carry confidence scores, experimental methods, context annotations, supporting PubMed identifiers, and provenance. Edge types are grounded in Relation Ontology properties, with more specific subtypes and explicit inverse relationships, and node types come from OBO ontologies and Rfam categories. This combination of standard identifiers, ontology-typed edges, and explicit edge properties is what lets queries filter by context and lets machine-learning models consume semantic content, not just topology.","core_discovery":"The discovery is a construction claim: 80 public repositories and 11 biomedical ontologies can be reconciled into a single property graph with 6,553,767 nodes and 99,936,712 relationships, and the reconciliation is what unlocks new analysis. RNA identities are grounded at isoform level—RNAcentral identifiers for non-coding RNAs, Ensembl transcript identifiers for mRNAs—so isoforms that differ by a few nucleotides are no longer collapsed. Relationship types come from the Relation Ontology at varying granularity, and each relationship may carry score, method, context, interactor, variants, the source list, and PubMed identifiers. The paper argues this representation is enough to support contex","pith_inferences":["Because context strings such as \"hela cell\" are stored as source-provided values rather than mapped to a controlled vocabulary, the same interaction context may be expressed in several spellings; normalizing context to ontology terms would make context-aware queries more reliable than the current examples suggest.","The time-stratified setup could be standardized as a benchmark for RNA link prediction: train on all edges with PubMed IDs up to year T, test on edges first reported after T, and report accuracy with the same unbiased pipeline, letting different embedding methods be compared on the same temporal split.","If the look-up tables and entity-linking rules are released as reusable artifacts, they could serve as a benchmark for identifier mapping in RNA databases; the paper reports no precision or recall for these mappings, so their accuracy is the main uncertainty."],"forward_implications":["Researchers can write Cypher queries that restrict interactions by experimental validation (e.g., western blotting), by sequence composition (e.g., uracil content above 25%), or by shared supporting literature across an miRNA–gene–disease triple.","The isoform-level identification scheme distinguishes transcript variants such as LINC-PINT-205 and LINC-PINT-206, which the earlier gene-level release collapsed into one Entrez identifier.","Context-aware pruning can reduce graph size by 19.8% in a sncRNA subgraph by collapsing only isomorphic groups whose members have high sequence alignment scores, preserving biologically meaningful differences.","Adding sequence and description embeddings to topological link prediction improves balanced accuracy in most tested cases, with gains up to +5 percentage points on RNA–gene prediction.","Time-stratified evaluation on the miRNAdisease view, trained on edges supported by pre-2022 papers, assigns plausibility scores above 0.5 to 82% of later-discovered interactions, suggesting the graph supports predictive use."],"supporting_citations":[{"why":"Defines the original RNA-KG construction and the look-up table strategy used to map proprietary identifiers to standard schemes.","marker":"[10]"},{"why":"Supplies RNAcentral identifiers used to ground non-coding RNA nodes at sequence level.","marker":"[60]"},{"why":"Supplies Ensembl transcript identifiers used to represent mRNA isoforms.","marker":"[19]"},{"why":"Supplies Rfam classification used to type and organize RNA nodes.","marker":"[32]"},{"why":"RNAInter is the source of the bulk of the graph's interacts-with edges.","marker":"[33]"},{"why":"Defines the unbiased link-prediction pipeline and experimental settings used for all reported accuracy numbers.","marker":"[65]"},{"why":"Provides node2vec embeddings used in homogeneous and multimodal link prediction.","marker":"[25]"},{"why":"Provides LINE embeddings used as a second homogeneous link-prediction baseline.","marker":"[62]"},{"why":"Provides TransE embeddings used for heterogeneous link prediction with multiple edge types.","marker":"[4]"},{"why":"Needleman–Wunsch global alignment is used to decide which isomorphic node groups can be safely collapsed in content-aware pruning.","marker":"[45]"}],"fun_headline_variants":["RNA-KG v2.0: 100M RNA links with tissue, cell, disease context","Context-aware RNA graph: 100M interactions, 91 sources","RNA interactome now searchable by tissue, cell line, disease","RNA-KG v2.0: every RNA interaction tagged with context and provenance"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The entire graph is only as sound as the look-up tables and entity-linking rules that map each source's proprietary identifiers to standard schemes (RNAcentral for non-coding RNAs, Ensembl for transcripts, and ontology terms), and the paper reports no measured accuracy for those mappings.","fun_headline_variants_meta":{"raw":{"variants":["RNA-KG v2.0: 100M RNA links with tissue, cell, disease context","Context-aware RNA graph: 100M interactions, 91 sources","RNA interactome now searchable by tissue, cell line, disease","RNA-KG v2.0: every RNA interaction tagged with context and provenance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1261,"prompt_tokens":711,"completion_tokens":550,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":465}},"tokens_in":455,"tokens_out":550,"duration_ms":6315,"temperature":1.0,"reasoning_tokens":465,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:06:08.029556+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Independently re-resolve a random sample of 1,000 RNA-KG v2.0 edges: take both endpoint identifiers back to the original source records and to the standard scheme (RNAcentral, Ensembl, or ontology term) via the published look-up tables, and count mismatches. A mismatch rate materially above zero, or a failure to reproduce the claimed 6,553,767 nodes and 99,936,712 relationships from the released code and input files, would falsify the construction claim.","supporting_citations":[{"cited_title":"An ontology-based knowledge graph for representing interactions involving rna molecul es","cited_arxiv_id":null,"evidence_quote":"Defines the original RNA-KG construction and the look-up table strategy used to map proprietary identifiers to standard schemes."},{"cited_title":"Rnacentral 2021: secondary structure integration, improved sequence search and new member databases","cited_arxiv_id":null,"evidence_quote":"Supplies RNAcentral identifiers used to ground non-coding RNA nodes at sequence level."},{"cited_title":"Ensembl 2025","cited_arxiv_id":null,"evidence_quote":"Supplies Ensembl transcript identifiers used to represent mRNA isoforms."},{"cited_title":"Rfam 14: expanded coverage of metagenomic, viral and microrna families","cited_arxiv_id":null,"evidence_quote":"Supplies Rfam classification used to type and organize RNA nodes."},{"cited_title":"Rnainter v4.0: Rna interactome repository with redefined confidence scoring system and improved accessibility","cited_arxiv_id":null,"evidence_quote":"RNAInter is the source of the bulk of the graph's interacts-with edges."},{"cited_title":"RNA Knowledge-Graph analysis through homogeneous embedding methods","cited_arxiv_id":null,"evidence_quote":"Defines the unbiased link-prediction pipeline and experimental settings used for all reported accuracy numbers."},{"cited_title":"Node2vec: Scalable feature learning for networks","cited_arxiv_id":null,"evidence_quote":"Provides node2vec embeddings used in homogeneous and multimodal link prediction."},{"cited_title":"LINE: Large-scale Information Network Embedding","cited_arxiv_id":null,"evidence_quote":"Provides LINE embeddings used as a second homogeneous link-prediction baseline."},{"cited_title":"Translating embeddings for modeling multi-relational data","cited_arxiv_id":null,"evidence_quote":"Provides TransE embeddings used for heterogeneous link prediction with multiple edge types."},{"cited_title":"Needleman and Christian D","cited_arxiv_id":null,"evidence_quote":"Needleman–Wunsch global alignment is used to decide which isomorphic node groups can be safely collapsed in content-aware pruning."}],"review_version":1}