GenBank2PubMed: Bridging Viral Genomic Data and the Scientific Literature with AI-Assisted Curation

GenBank2PubMed: Bridging Viral Genomic Data and the Scientific Literature with AI-Assisted Curation | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article GenBank2PubMed: Bridging Viral Genomic Data and the Scientific Literature with AI-Assisted Curation Kaiming Tao, Jinru Zhou, Yimam Getaneh, Robert W. Shafer This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6710551/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 22 Nov, 2025 Read the published version in Scientific Reports → Version 1 posted 11 You are reading this latest preprint version Abstract Background : GenBank entries of pathogenetic viral sequences are typically annotated with host species and epidemiological metadata. However, linking these entries to their corresponding published studies remains labor-intensive. Methods : We developed GenBank2PubMed , a computation pipeline that integrates GenBank sequence data with metadata from published studies. The pipeline aggregates GenBank entries into submission sets based on shared authorship, title similarity, submission dates, and the sequential nature of their accession numbers. Using automated methods, including GPT-4, we linked these submission sets to relevant publications – a challenging task given that many GenBank entries lack citation references. The result is a database in which viral sequences are annotated by host, country, and year of isolation. We also conducted a systematic review to assess how frequently published studies reporting sequences included GenBank submissions. We applied GenBank2PubMed to three high-mortality viruses with outbreak potential: Crimean-Congo Hemorrhagic Fever (CCHF) virus, Lassa virus, and Nipah virus. Results : We identified 193 CCHF virus submission sets (4,754 entries), 78 Lassa virus sets (2,663 entries), and 34 Nipah virus sets (355 entries). Of these, 173 (CCHF), 64 (Lassa), and 31 (Nipah) were linked to published studies. Integration with publication data enriched the contextual and epidemiological metadata for each set. Additionally, our literature review found that 80.1% of CCHF, 86.6% of Lassa, and 87.5% of Nipah virus studies reporting sequences had corresponding GenBank submissions. GenBank submission sets and relational databases for each virus are available at https://hivdb.stanford.edu/genbank2pubmed/; the pipeline is available at https://github.com/hivdb/GenBankRefs. Conclusions : Creating submission sets facilitates the organization of GenBank data into browsable spreadsheets and queryable databases. GPT-4 contributed to linking GenBank entries with published studies and extracting metadata, although manual validation remained essential for accuracy. GenBank2PubMed represents a significant step toward integrating GenBank viral sequences with the scientific literature in which they are reported. Health sciences/Diseases/Infectious diseases/Viral infection Biological sciences/Computational biology and bioinformatics/Databases/Genetic databases Biological sciences/Computational biology and bioinformatics/Data acquisition Biological sciences/Computational biology and bioinformatics/Data integration Biological sciences/Computational biology and bioinformatics/Literature mining Biological sciences/Computational biology and bioinformatics/Sequence annotation Virus sequences molecular epidemiology phylogenetics GenBank PubMed GPT-4o Crimean Congo Hemorrhagic Fever virus Lassa virus Nipah virus Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 INTRODUCTION The National Center for Biotechnology Information (NCBI) GenBank, DNA Databank of Japan (DDBJ), and the European Nucleotide Archive (ENA) are the three major public nucleotide sequence databases comprising the International Nucleotide Sequence Database Collaboration (INSDC) ( 1 ). These three databases—hereafter collectively referred to as GenBank—share data daily to ensure that their contents remain synchronized. Viral sequences in these databases are typically annotated with information about the host species, the specimen types from which the viruses were isolated, and, for human-infecting viruses, demographic data about the infected individuals. For treatable pathogenic human viruses, data on antiviral treatments received by individuals from whom the viruses were isolated are sometimes, but rarely, included. For the past 25 years, we have maintained the Stanford HIV Drug Resistance Database (Stanford HIVDB), a curated database containing HIV sequences annotated with antiviral treatment histories of the individuals from whom the viruses were obtained ( 2 ). To support this effort, we developed manual and computational approaches to annotate sequence data from GenBank with data extracted from research publications. Our software pipeline aggregates individual GenBank entries into submission sets likely to be linked based on the similarity of their authors, titles, and submission dates, as well as the sequential nature of their accession numbers. We have also long been interested in the extent to which published studies reporting virus sequences submit those sequences and their associated metadata to GenBank ( 3 , 4 ). More recently, we have explored the use of automated methods—including large language models (LLMs)—to integrate data from published studies with the sequence data in GenBank ( 5 , 6 ). In this study, we formalized and generalized the software pipeline originally developed for the Stanford HIVDB so that it could be applied to other pathogenic human viruses. We also developed additional automated tools, including LLMs, to link GenBank submission sets with their corresponding published studies – a non-trivial task, given that many GenBank entries lack publication references. The result is a software toolkit that we call GenBank2PubMed capable of rapidly generating a database in which all available gene sequences of a virus are annotated by their host, country and year of isolation. For this study, we selected three high-mortality viruses with outbreak potential – Crimean Congo Hemorrhagic Fever Virus, Lass Virus, and Nipah Virus ( 7 ). In parallel, we conducted a systematic review of the literature to assess the completeness of our automated methods at identifying published studies that were not included in the GenBank records. This manual review allowed us to determine how often sequences from published studies were submitted to GenBank. METHODS Dictionary of essential virus data For each virus we created a dictionary of essential virus data including (i) the formal species name for the virus used by the NCBI; (ii) a list of each virus’s gene names and their synonyms; (iii) the NCBI virus reference sequence and its GenBank accession number; (iv) the formal names of virus host species mapped to commonly used names. For example, Homo sapiens was mapped to “humans”; various tick species were mapped to “ticks”; and bat species were mapped to “bats”. Defining GenBank submission sets We searched GenBank using the following formal virus species names: Orthonairovirus haemorrhagiae for Crimean Congo Hemorrhagic Fever (CCHF) virus, Mammarenavirus lassaense for Lassa virus, and Henipavirus nipahense for Nipah virus. These searches yielded composite GenBank files containing the complete set of GenBank entries for each virus. The composite files were processed using the BioPython library to extract the GenBank reference, feature, descriptive, and sequence data for each entry. Individual GenBank entries with the same PubMed ID (PMID) or highly similar study titles, lists of authors, submission years, and sequential accession numbers were aggregated into GenBank submission sets, each containing one or more GenBank entries. Submission sets containing artificial constructs, patented sequences, unverified sequences, or laboratory manipulated viruses were identified by their GenBank annotation and excluded from further analysis. Linking GenBank submission sets to published studies GenBank submission sets containing a PMID were directly linked to the corresponding published study. For submission sets without a PMID, we employed two automated methods to identify the corresponding published study: (i) querying the PubMed API for publications authored by all pairs of contributors of the GenBank submission set and (ii) using the GPT-4o API to search the web for potential publications associated with the submission (Supplementary Text File). We then assessed whether these methods converged on the same published study. To confirm linkage, we retrieved the publication’s PDF and checked for matching GenBank accession numbers. If no accession numbers were present, we evaluated the study’s title, abstract, and author list to determine the likelihood of its association with the GenBank submission set. The software pipeline for aggregating GenBank entries into submission sets and to integrate the submission sets with the appropriate published studies is summarized in Fig. 1 . Confirmatory PubMed search and literature review To evaluate the performance of our automated methods in linking published studies to GenBank submission sets, we conducted a manual PubMed search and literature review for each virus. We used the search terms “Crimean-Congo Hemorrhagic Fever Virus”, Lassa Virus”, and “Nipah Virus”. Two reviewers screened all titles and abstracts to identify publications potentially reporting viral sequence data. Flagged studies underwent a full-text review to confirm whether it indeed reported viral sequences. The manual PubMed search served an additional purpose. It enabled us to determine what proportion of published studies reporting viral sequences submitted the sequences to GenBank. GenBank submission sets that could not be linked to a published study using either of our two automated methods or manual PubMed review were further investigated to determine if they had been published in a non-PubMed-indexed journal. For these cases, we searched SCOPUS using the virus name along with one or more of the submission set authors. We then reviewed the resulting publications to assess whether they referenced the same accession numbers or if the similarity in author lists and publication titles indicated that they were associated with the GenBank submission set. Creating a relational database We created SQLite relational databases for each virus, each containing four core tables (Fig. 2 ): (i) tblGBEntries, which contains GenBank accession numbers and metadata associated with each GenBank entry; (ii) tblSequences, which contains separate rows for each gene in a GenBank entry and that entry’s nucleotide and amino acid sequences. The start and stop position of the amino acid sequence following alignment to the virus’s reference sequence for that gene and it’s percent identity to the reference sequence are included; (iii) tblGBSubmissionSetRefs, which contains reference data for each GenBank submission set; and (iv) tblPublications, which contains reference data for each published study. The database also contains six supporting tables (Fig. 2 ): (i) tblGBRefLink, which manages the many-to-many relationship between tblGBSubmissionSetRefs and tblGBEntries; (ii) tblGBPubRefLink, which manages the occasional many-to-many relationship between tblGBSubmissionSets and tblPublications; (iii) tblPublicationsData, which stores the raw data extracted from publications; (iv) tblInsertions, which stores the insertions within each sequence relative to the virus’s reference genes; (v) tblQA, which stores the number of unsequenced bases, number of frameshifts, number of indels, number of stop codons, and number of differences from the reference sequence; and (vi) tblIsolates, which links virus isolates having more than one accession number, a situation that arises when authors create separate GenBank submissions for different genes of the same virus. Phylogenetic analyses To evaluate the consistency of our database with published phylogenetic analyses, we queried the database for near-complete gene sequences along with associated metadata including host, country, and sample year, and performed our own phylogenetic analyses. Each sequence was aligned to the reference nucleic acid sequence using Scikit-Bio ( https://scikit.bio/ ). Insertions were excluded from the alignment, while deletions and ambiguous characters were retained. To limit the CCHF and Lassa virus trees to 100 sequences that represent the diversity of these virus populations, we used the pplacer program to select those sequences that minimize the average distance to the closest leaf ( 8 ). Phylogenetic trees were constructed using IQTree2 with a general time reversible + gamma distribution model ( 9 ). Data extraction from published studies For each English-language published study determined to be reporting viral sequences, we extracted seven key data elements: (i) the host species from which samples were obtained; (ii) the type of specimen sequenced; (iii) the country or countries of the sequenced samples; (iv) the years in which virus samples were collected; (v) the viral genes sequenced; (vi) the method of sequencing; and (vii) the reported GenBank accession numbers. Data extraction was performed both manually and using GPT-4o. GPT-4o’s accuracy was defined as the frequency with which its extracted data matched the manual review. GPT-4o’s data extractions that did not exactly match the manual review were classified as either partially or completely incorrect. For example, a result was considered partially incorrect if GPT-4o identified the correct country but omitted another relevant country. However, if GPT-4o identified an incorrect country or misidentified the country entirely, the result was classified as completely incorrect. RESULTS CCHF virus GenBank submission sets and linkage to published studies As of November 27, 2024, there were 4,848 GenBank entries. 94 entries were excluded because they represented artificial constructs, patented sequences, unverified sequences, or laboratory viruses. The remaining entries were aggregated into 193 submission sets containing 4,754 entries. Of the 193 submission sets, 173 (89.6%) were linked to a publication including 115 that contained a PMID (Fig. 3 A). Of the 78 submission sets without a PMID, 48 were linked to a published study through at least one of two automated methods: PubMed author query (n = 27) and GPT-4o web search (n = 39). An additional 9 submission sets were only identified by our manual PubMed search and literature review. One submission set that could not be linked to a published study by any other method was linked to a study by searching SCOPUS. The 20 remaining submission sets that could not be linked to a published study contained 147 GenBank entries. Twelve had the title “Direct submission,” making it difficult to identify a corresponding publication. Four were submitted to GenBank between 2023 and 2024. Supplementary Table 1 provides details on the 78 CCHF virus GenBank submission sets without a linked PMID in GenBank and the methods by which they were linked to a published study. Relational database Of the GenBank sequences, 3,417 (71.8%) were obtained from human samples, 990 (20.8%) from ticks, and 2.2% from livestock. For 5.1% of sequences, the host was not indicated in the GenBank record. All samples from humans in which the specimen type was provided were from blood, serum, or plasma. The most common countries from which samples were obtained were Russia (46.8%), Turkey (14.3%), Pakistan (5.8%), India (4.7%), Kosovo (4.5%), Iran (3.7%), China (3.5%), South Africa (2.7%), Uganda (1.4%), Spain (1.1%), Kazakhstan (1.1%), and Bulgaria (1.0%). Thirty-five countries, each reporting less than 1% of samples, comprised 7.8% of GenBank sequences. For 1.6% of sequences no country could be identified. The sample year was available for 92.7% of sequences. The distribution of samples by decade was as follows: 1950–1959 (0.2%), 1960–1969 (1.3%), 1970–1979 (1.2%), 1980–1989 (2.9%), 1990–1999 (0.7%), 2000–2009 (17.6%), 2010–2019 (69.2%), and 2020–2024 (7.1%). The full genome comprising each of the three CCHF virus segments – Large (L), Medium (M), and Small (S) – was available for 4.9% of virus isolates. For 92.5% of isolates, a single segment was reported and for 2.6% of isolates, two segments were reported. Overall, the part of the S segment containing the nucleoprotein gene was reported for 52.4% of isolates while approximately 32.7% of isolates reported M segments and 27.3% reported the L segment. Phylogenetic analyses Figure 4 shows the phylogenetic tree containing 100 near-complete L segments encompassing 10500 nucleotides. It demonstrates each of the six previously reported lineages ( 10 – 13 ). Although each lineage tends to occur in specific geographic regions, there are some exceptions postulated to arise from long distance virus transfers through migratory birds ( 14 ). Lassa virus GenBank submission sets and linkage to published studies As of November 27, 2024, there were 2,780 GenBank entries. We excluded 117 entries because they represented artificial constructs, patented sequences, unverified sequences, or laboratory isolates. The remaining entries were aggregated into 78 submission sets containing 2,663 entries. Of the 78 submission sets, 64 (80.0%) were linked to a published study including 44 submission sets containing a PMID (Fig. 3 B). Of the 34 submission sets that did not contain a PMID, 21 were linked to a published study through at least one of two automated methods: PubMed author query (n = 18) and GPT-4o web search (n = 20). No additional submission set was identified by our manual searches of PubMed and SCOPUS. The 13 submission sets that could not be linked to a publication contained 171 entries. Six sets had the title “Direct submission” making it difficult to identify a corresponding published study. Two sets were submitted to GenBank between 2023 and 2024. Supplementary Table 2 provides details on the 34 Lassa virus GenBank submission sets without a linked PMID and the methods by which they were linked to a published study. Relational database Of the 2,663 sequences in GenBank, 56.5% were from human samples, 37.9% were from rodents, and 0.3% were from other animals. For 5.3% the host was not known. For the 921 human samples for which the specimen type was provided, 98.8% were from blood, 0.7% from CSF, 0.2% from urine, 0.2% from pleural fluid, and 0.1% from breast milk. The countries from which samples were obtained were Nigeria (54.3%), Guinea (18.1%), Sierra Leone (15.1%), Liberia (3.1%), Cote d'Ivoire (2.2%), Benin (1.2%), and Mali (1.0%). For 4.7% of samples, no country was identified. The sample year was available for 94.0% of sequences. The distribution of samples by decade was as follows: 1960–1969 (0.1%), 1970–1979 (0.7%), 1980–1989 (0.8%), 1990–1999 (0.1%), 2000–2009 (16.7%), 2010–2019 (78.9%), and 2020–2024 (2.7%). The full genome of all four of the virus’ proteins were available for 11.0% of isolates. For 50.8% of isolates a single gene sequence was available and for 38.2% of isolates, two or three gene sequences were available. Nucleocapsid gene sequences were available for 55.7% of sequences; G gene sequences for 55.1% of isolates; L gene sequences for 33.6% of isolates; and Z gene sequences for 27.3% of isolates. Phylogenetic analyses Figure 5 shows the phylogenetic tree containing 100 near-complete L genes encompassing 4600 nucleotides. It reveals four distinct clusters from Nigeria, which have been designated lineages I to IV, a cluster from Mali designated lineage V, and several smaller clusters from Liberia, Sierra Leone, and Guinea, all classified under lineage IV ( 15 , 16 ). Additionally, a cluster from Togo and Benin is identified as lineage VII ( 15 , 16 ). Several of these lineages include isolates from both humans and rodents, with sequences clustering primarily by country. Nipah virus GenBank submission sets and linkage to published studies As of November 27, 2024, there were 407 GenBank entries. 52 entries were excluded because they represented artificial constructs, patented sequences, unverified sequences, or laboratory viruses. The remaining entries were aggregated into 34 submission sets containing 355 entries. Of the 34 submission sets, 31 (91.2%) were linked to a publication including 22 submission sets containing a PMID (Fig. 3 C). Of the 12 submission sets without a PMID, 9 were linked to a published study through at least one of two automated methods: PubMed author query (n = 4) and GPT-4o web search (n = 9). No additional submission set was identified by our manual searches of PubMed and SCOPUS. Three submission sets containing 33 GenBank entries could not be linked to a published study. One had the title “Direct submission” making it difficult to identify a corresponding published study. One set was submitted in 2024. Supplementary Table 3 provides details on the 12 Nipah virus GenBank submission sets without a linked PMID and the methods by which they were linked to a published study. Relational database Among the 355 sequences in GenBank, 62.5% were from bats, 32.7% were from humans, 1.4% from pigs, and 0.3% from dogs. For 3.1% (n = 11) the host could not be determined. For the 134 human samples for which the specimen type was provided, 48.6% were from throat swabs or saliva, 45.5% were from urine, 3.7% were from blood, 1.5% were from CSF, and 0.7% were from breast milk. The countries from which human and bat samples were obtained were from Bangladesh (36.1%), Thailand (29.0%), India (17.7%), Cambodia (9.0%), Malaysia (4.5%), Indonesia (1.4%), and Sri Lanka (1.1%). For 1.1% of sequences the country could not be determined. The sample year was available for 91.3% of GenBank submissions: 1998–1999 (1.9%), 2000–2009 (28.1%), 2010–2019 (57.1%), 2020–2024 (13.0%). The full genome containing each of the six Nipah virus proteins was available for 31.5% of isolates. For 61.4% of isolates, just the nucleocapsid gene was sequenced. For the remaining 7.1% of sequences one or more genes other than nucleocapsid was sequenced including G, M, P, L, or F. Phylogenetic analyses Figure 6 shows the phylogenetic tree containing 79 complete N genes encompassing 1596 nucleotides. It demonstrates that sequences from Bangladesh and India comprise one major lineage while sequences from Malaysia form a second major lineage. The lineage containing sequences from Bangladesh and India contains two sublineages from Bangladesh and one sublineage from India. Sequences belonging to different hosts (i.e., human and bat) cluster by country rather than host. Sequences from bats in Thailand and Cambodia clustered with the Malaysian sequences. Confirmatory PubMed search and literature review CCHF virus A PubMed search using the term “Crimean Congo Hemorrhagic Fever virus” performed December 10, 2025, retrieved 2,131 records. After reviewing titles and abstracts, 261 publications underwent full-text review and 181 were confirmed to be reporting CCHF virus sequences. Among these, 145 (80.1%) were successfully matched to a GenBank submission set. As noted above, literature searches identified ten additional matching studies that could not be linked through automated methods alone (nine via the PubMed search and one via the SCOPUS search). Of the remaining 36 studies (19.9%) that could not be matched to a GenBank submission set, two contained sequences that were submitted to the NCBI Short Read Archive (SRA), two reported submitting their sequences but did not provide accession numbers, and two reported that their sequences were available upon request. The 36 CCHF virus PubMed publications without GenBank-linked sequences is provided in Supplementary Table 4. Lassa virus A PubMed search using the term “Lassa virus” performed January 6, 2025, retrieved 1,680 records. After reviewing titles and abstracts, 103 publications underwent full-text review and 67 were confirmed to be reporting Lassa virus sequences. Among these, 58 (86.6%) were successfully matched to a GenBank submission set. The PubMed search did identified one additional study that could not be linked through automated methods alone. Of the remaining nine studies (13.4%) that could not be matched to a GenBank submission set, two were case reports. The nine Lassa virus PubMed publications without GenBank-linked sequences is provided in Supplementary Table 5. Nipah virus A PubMed search using the term “Nipah virus” performed December 15, 2024, retrieved 1,515 records. After reviewing titles and abstracts, 76 publications underwent full-text review and 32 were confirmed to be reporting Nipah virus sequences. Among these, 28 (87.5%) were successfully matched to a GenBank submission set. The PubMed search did not identify additional studies that could not be linked through automated methods alone. Of the remaining four studies (12.5%) that could not be matched to a GenBank submission set, one reported three sequences and stated that the sequences had been submitted to GenBank but did not provide accession numbers. The four Nipah virus PubMed publications without GenBank-linked sequences are listed in Supplementary Table 6. Accuracy of GPT-4 data extraction For each of the published studies, we assessed the accuracy of GPT-4 at extracting information on the hosts, specimen types, sequencing method, sample years, countries, genes, and GenBank accession numbers. Figure 7 shows GPT-4 accuracy for each data type for each virus. For the 223 CCHF virus, 72 Lassa virus, and 37 Nipah virus studies, the overall accuracy of GPT-4 was 91.2%, 92.3%, and 89.6%, respectively. An additional 2.6%, 1.8%, and 3.5% of questions for each virus were partially correct. The data extracted from published studies modestly reduced the proportion of GenBank entries for which the host, country, or sample year was unknown. The proportion of entries for which the host species, country, or sample year were missing for the three viruses combined based on the GenBank record alone was 4.3%, 5.3%, and 4.3%, respectively. Following extraction of data from published papers, these proportions were 0.6%, 0.7%, and 0.1%, respectively. Resources We created a dedicated website for this project, which provides comprehensive information for each virus: https://hivdb.stanford.edu/genbank2pubmed/ . The website includes: (i) A link to download the virus’s SQLite database; (ii) A link to download Excel files summarizing each virus’s GenBank submission sets; (iii) A dynamic table summarizing metadata for each virus’s GenBank submission set, including virus host, country, specimen, year, and sequenced genes; (iv) A static list of GenBank submission sets linked to a published study; (v) A static list of those GenBank submission sets not linked to a published study; (vi) A static list of published studies reporting sequences not found in GenBank; (vii) Phylogenetic analyses of viral gene sequences. DISCUSSION We developed a software pipeline called GenBank2PubMed that created annotated relational databases for three high-mortality viruses with outbreak potential. The pipeline consists of the following steps. First, Python scripts group GenBank entries into submission sets based on highly similar titles, author lists, submission years, and sequential accession numbers. Only those GenBank entries that represent clinical or zoonotic viruses were included while patented sequences, viral constructs, and laboratory isolates were excluded. Second, two automated methods identify published studies associated with those submission sets for which the GenBank entries lack PMIDs. Third a manual PubMed search is performed to identify studies that were not identified by our automated methods and to obtain metadata that was not already contained in GenBank. Fourth, data from GenBank entries and linked published studies are integrated into a relational database. To create the GenBank submission sets, we processed GenBank files containing all virus entries, ranging in size from 4 mb for Nipah virus to 25 mb for CCHF virus, converting them into Excel files ranging in size from 17 kb to 68 kb that can be browsed by reference data, sequenced gene, host, country, and year. Across all three viruses, we identified a total of 305 GenBank submission sets. Of these, 181 (59.3%) included a PMID in their GenBank entries, while 124 (40.7%) lacked any link to a published study. For these 124 sets, automated PubMed searches and GPT-4 web searches combined to identify 78 (62.9%) linked published studies. Manual literature searches yielded just 10 (8.1%) linked studies that were not found by automated methods. The remaining 36 submission sets – 29.0% of those lacking PMIDs and 11.8% of the total – could not be associated with a publication; among these were 19 labeled “Direct submission” and seven submitted to GenBank between 2023 and 2024. The relational databases created are useful to researchers determining the extent of virus sequencing by host, country, sample year, and gene. This may influence the design of surveillance studies to address gaps in publicly available virus sequence data. It also facilitates the selection of sequences for phylogenetic analyses. Because the sequences are stored in a virtual alignment to established reference sequences for each virus gene, it is also possible to identify those sequences encompassing sufficient coverage for phylogenetic analysis. The database also contains quality control data for each sequence such as the number of indels, stop codons, unsequenced bases, and frameshifts which can be used to exclude certain sequences from analysis. Although the manual PubMed search identified only ten additional publications not captured by automated methods, it allowed us to address a related question: how frequently do authors who report viral sequences in their publications also submit those sequences to GenBank? We found that 80.1% of CCHF virus studies, 86.6% of Lassa virus studies, and 87.5% of Nipah virus studies that reported sequences had corresponding submissions in GenBank. This high rate of GenBank submission likely reflects the relatively limited number of sequences available for these viruses, in contrast to more extensively sequenced pathogens. For comparison, we previously reported that approximately 60% of 934 HIV drug resistance studies published between 2010 and 2019 had submitted sequence data to GenBank ( 3 ). In recent years, there have been multiple studies investigating the application of LLMs to a range of tasks related to the processing of biomedical literature, including publication triage, content summarization, and data extraction ( 17 – 19 ). In this study, we used GPT-4o to identify published studies linked to GenBank submission sets and to extract data from published studies. We found that its accuracy ranged from about 80–95% depending on the type of data extracted. We expect that this accuracy will improve over time as LLMs improve and potentially with fine-tuning, a process that entails the continued training of a pre-trained LLM on domain-specific material for the purpose of adapting the model to specific tasks ( 20 , 21 ). Indeed, we have recently reported that the accuracy of GPT-4-mini and the open source model Llama 3.1 Instruct 70B at answering questions about studies on HIV drug resistance can be significantly improved by fine-tuning ( 6 ) Limitations First, because LLMs can be inaccurate, their findings required manual verification. Published studies linked to GenBank submission sets by automated methods had to be manually reviewed if they lacked matching accession numbers. Metadata extracted from publications also needed to be manually validated for accuracy. Second, 11.8% (36/305) of GenBank submission sets could not be linked to a published study using any approach, including two automated methods, a manual PubMed search, and a final SCOPUS search. These sequences may have been described in studies we were unable to identify. Finally, although identifying the corresponding studies provided useful context, the amount of additional epidemiological data extracted was smaller than expected; the proportion of sequences lacking such data declined only modestly – from approximately 5–1% – after data extraction. Conclusions and future directions The creation of GenBank submission sets makes it possible for the large number of GenBank entries for a virus to be consolidated into a tractable number of submission sets containing just those viruses of epidemiological relevance. Automated methods including an LLM can identify published studies that describe the GenBank submission sets in more detail than indicated by the GenBank entries alone. Finally, as LLMs improve at performing systematic reviews – specifically identifying published papers and extracting specific data – the process of creating virus sequence databases that include both data from GenBank and from published studies will be streamlined. Abbreviations CCHF Crimean-Congo Hemorrhagic Fever NCBI National Center for Biotechnology Information DDBJ DNA Databank of Japan ENA European Nucleotide Archive INSDC International Nucleotide Sequence Database Collaboration LLM Large language model PMID PubMed ID SRA Short Read Archive Declarations Ethics approval and consent to participate Not applicable. This study only utilized publicly available data. Consent for publication Not applicable. This study only utilized publicly available data. Competing interests The authors declare no competing interests. Funding The research reported in this paper was supported by a grant from the NIH: NIH/NIAID R24AI13661806. Author Contribution KT and RWS conceived of the project; KT and JZ wrote the software for the project; KT, JZ, YM, and RWS performed literature reviews; KT, JZ, YM, and RWS performed data analysis; KT, JZ, YM, and RWS created one or more figures; RWS wrote the manuscript; All authors reviewed the manuscript. Acknowledgements Not applicable Data Availability The data in this study are available online (https://hivdb.stanford.edu/genbank2pubmed/), in a GitHub repository (https://github.com/hivdb/GenBankRefs), and in six supplementary files. References Karsch-Mizrachi, I., Takagi, T. & Cochrane, G. The international nucleotide sequence database collaboration. Nucleic Acids Res. 46 (Database issue), D48–51 (2018). Shafer, R. W., Jung, D. R., Betts, B. J., Xi, Y. & Gonzales, M. J. Human Immunodeficiency Virus Reverse Transcriptase and Protease Sequence Database. Nucleic Acids Res. 28 (1), 346–348 (2000). Rhee, S. Y. et al. Public availability of HIV-1 drug resistance sequence and treatment data: a systematic review. Lancet Microbe . 3 (5), e392–e398 (2022). Inzaule, S. C. et al. Recommendations on data sharing in HIV drug resistance research. PLoS Med. 20 (9), e1004293 (2023). Tao, K. et al. GPT-4 performance on querying scientific publications: reproducibility, accuracy, and impact of an instruction sheet. BMC Med. Res. Methodol. 24 (1), 139 (2024). Tao, K. et al. Fine-tuned large language models for answering questions about full-text biomedical research studies [Internet]. medRxiv; 2024 [cited 2025 Mar 23]. p. 2024.10.28.24316263. Available from: https://www.medrxiv.org/content/ 10.1101/2024.10.28.24316263v2 Mehand, M. S., Al-Shorbaji, F., Millett, P. & Murgue, B. The WHO R&D Blueprint: 2018 review of emerging infectious diseases requiring urgent research and development efforts. Antiviral Res. 159 , 63–67 (2018). Matsen, F. A. I. V., Gallagher, A. & McCoy, C. O. Minimizing the Average Distance to a Closest Leaf in a Phylogenetic Tree. Syst. Biol. 62 (6), 824–836 (2013). Minh, B. Q. et al. IQ-TREE 2: New Models and Efficient Methods for Phylogenetic Inference in the Genomic Era. Mol. Biol. Evol. 37 (5), 1530–1534 (2020). Carroll, S. A., Bird, B. H., Rollin, P. E. & Nichol, S. T. Ancient common ancestry of Crimean-Congo hemorrhagic fever virus. Mol. Phylogenet. Evol. 55 (3), 1103–1110 (2010). Bente, D. A. et al. Crimean-Congo hemorrhagic fever: History, epidemiology, pathogenesis, clinical syndrome and genetic diversity. Antiviral Res. 100 (1), 159–189 (2013). Lukashev, A. N. et al. Phylogeography of Crimean Congo Hemorrhagic Fever Virus. PLOS ONE . 11 (11), e0166744 (2016). Shahhosseini, N. et al. Crimean-Congo Hemorrhagic Fever Virus in Asia, Africa and Europe. Microorganisms 9 (9), 1907 (2021). Lindeborg, M. et al. Migratory Birds, Ticks, and Crimean-Congo Hemorrhagic Fever Virus - Volume 18, Number 12—December 2012 - Emerging Infectious Diseases journal - CDC. [cited 2025 Apr 2]; Available from: https://wwwnc.cdc.gov/eid/article/18/12/12-0718_article Forni, D. & Sironi, M. Population structure of Lassa Mammarenavirus in West Africa. Viruses 12 (4), 437 (2020). Garry, R. F. Lassa fever — the road ahead. Nat. Rev. Microbiol. 21 (2), 87–96 (2023). Dennstädt, F., Zink, J., Putora, P. M., Hastings, J. & Cihoric, N. Title and abstract screening for literature reviews using large language models: an exploratory study in the biomedical domain. Syst. Rev. 13 (1), 158 (2024). Lieberum, J. L. et al. Large language models for conducting systematic reviews: on the rise, but not yet ready for use—a scoping review. Journal of Clinical Epidemiology [Internet]. 2025 May 1 [cited 2025 Mar 28];181. Available from: https://www.jclinepi.com/article/S0895-4356(25)00079-4/fulltext Santos, Á. O., da Silva, E. S., Couto, L. M., Reis, G. V. L. & Belo, V. S. The use of artificial intelligence for automating or semi-automating biomedical literature analyses: A scoping review. J. Biomed. Inform. 142 , 104389 (2023). Hu, E. J. et al. LoRA: Low-Rank Adaptation of Large Language Models [Internet]. arXiv; [cited 2024 Jul 5]. (2021). Available from: http://arxiv.org/abs/2106.09685 Tinn, R. et al. Fine-tuning large neural language models for biomedical natural language processing. Patterns 4 (4), 100729 (2023). Additional Declarations No competing interests reported. Supplementary Files SupplementaryFiles.zip Cite Share Download PDF Status: Published Journal Publication published 22 Nov, 2025 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 23 Jun, 2025 Reviews received at journal 16 Jun, 2025 Reviews received at journal 11 Jun, 2025 Reviewers agreed at journal 04 Jun, 2025 Reviewers agreed at journal 03 Jun, 2025 Reviewers agreed at journal 28 May, 2025 Reviewers invited by journal 28 May, 2025 Editor invited by journal 27 May, 2025 Editor assigned by journal 22 May, 2025 Submission checks completed at journal 21 May, 2025 First submitted to journal 20 May, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6710551","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":463814922,"identity":"2e6598dd-58a3-4fce-b1f2-c10c629c425a","order_by":0,"name":"Kaiming Tao","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Kaiming","middleName":"","lastName":"Tao","suffix":""},{"id":463814923,"identity":"051e63bd-4ab6-48d2-bb26-5930237911ca","order_by":1,"name":"Jinru Zhou","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Jinru","middleName":"","lastName":"Zhou","suffix":""},{"id":463814924,"identity":"4345f249-4e7b-426b-ab3c-066259cb8137","order_by":2,"name":"Yimam Getaneh","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Yimam","middleName":"","lastName":"Getaneh","suffix":""},{"id":463814925,"identity":"e258a755-cc10-4d4e-b85c-ed943acd6f7b","order_by":3,"name":"Robert W. Shafer","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA2klEQVRIie2PMQuCQBSAX4suB60HZf4FQ1B/zkmQi7YI4tBwILj6E/wLtjifBNdiu1M0NTXUEk3RWWOhtjXcN33D+x7vAUgkfwkCiAFmLxcyogBGf1IDmKDAS4YnLh2c2Nm+urD44OV5cjoK0aialLgrmTarBWZ1GBRcsQ0hJkU86kwwRsbkmpKgUMBqxaXYt7qTcW3eqwfx9FS94apN9HNPAmJnRQkBjqx3glFPIs5wGCfzgvuReIqYKVqGTt9hDVsTXU92JWYx0TJ1u2m6kk+U38YlEolE8o0nTRZLj0hFMEUAAAAASUVORK5CYII=","orcid":"","institution":"Stanford University","correspondingAuthor":true,"prefix":"","firstName":"Robert","middleName":"W.","lastName":"Shafer","suffix":""}],"badges":[],"createdAt":"2025-05-20 19:38:09","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6710551/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6710551/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-025-28386-8","type":"published","date":"2025-11-22T15:57:09+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":83666203,"identity":"2b814155-4b15-4cba-b71a-ae47be69914c","added_by":"auto","created_at":"2025-05-30 11:42:01","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1685035,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSoftware pipeline for integrating GenBank virus entries with published literature\u003c/strong\u003e. The pipeline begins by aggregating GenBank entries into submission sets, excluding entries that are not epidemiologically relevant, and using automated methods to identify linked published studies. In the second step, manual PubMed searches are performed to identify additional publications not captured by automated methods. Each virus gene sequence is aligned to its corresponding NCBI reference sequence, allowing the creation of a virtual alignment across sequences. The resulting virus sequences and associated metadata are organized in a relational database, which is accessible via a publicly available website that also offers downloadable Excel files containing the submission sets. Phylogenetic trees were generated to validate the consistency of the curated sequences with previously published studies.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-6710551/v1/49ecb440455b052602c0573e.png"},{"id":83666195,"identity":"4414557e-f0a0-44ab-9272-3e36bacf1443","added_by":"auto","created_at":"2025-05-30 11:41:46","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":218322,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSchema for each virus database\u003c/strong\u003e. Each database contains four core tables and five supporting tables. The core tables are: (i) tblGBEntries – Contains GenBank accession numbers and metadata for each entry; (ii) tblSequences – Stores individual gene sequences for each entry; (iii) tblGBSubmissionSetRefs – Contains reference data for each submission set; (iv) tblPublications – Contains citation data for published studies. The supporting tables are: (i) tblGBRefLink – Manages many-to-many relationship between tblGBEntries and tblGBSubmissionSetRefs; (ii) tblGBPubRefLink – Handles occasional many-to-many relationships between tblGBSubmissionSetRefs and tblPublications; (iii) tblPublicationsData – Stores raw data extracted from publications; (iv) tblInsertions – Stores the insertions within each sequence relative to the virus’s reference genes; (v) tblQA – Stores the number of unsequenced bases, number of frameshifts, number of indels, number of stop codons, and number of differences from the reference sequence; and (vi) tblIsolates – Links virus isolates having more than one accession number, a situation that arises when separate GenBank submissions are made for different genes of the same virus.\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-6710551/v1/93add098adb54a91b04fab08.png"},{"id":83666190,"identity":"19e77574-1860-4682-81f8-4076ba500f28","added_by":"auto","created_at":"2025-05-30 11:41:46","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":323550,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eIdentification of published studies associated with each GenBank submission sets\u003c/strong\u003e. \u003cstrong\u003eA. CCHF virus\u003c/strong\u003e – 193 submission sets including 78 lacking a PubMed ID (PMID). Of these, 48 (61.5%) linked studies were identified by automated methods and 10 (12.8%) only by a manual systematic literature review; 20 (25.6%) could not be linked to a published study. \u003cstrong\u003eB. Lassa virus\u003c/strong\u003e– 78 submission sets including 34 lacking a PMID. Of these, 21 (61.8%) linked studies were identified by automated methods. No additional studies were identified by manual systematic literature review. 13 (38.2%) could not be linked to a published study. \u003cstrong\u003eC. Nipah virus\u003c/strong\u003e – 34 submission sets including 12 lacking a PMID: 9 (75.0%) linked studies were identified by automated methods. No additional studies were identified by manual systematic literature review. 3 (25.0%) could not be linked to a published study. D. \u003cstrong\u003eCombined results\u003c/strong\u003e – 305 submission sets including 124 lacking a PMID. Of these, 78 (62.9%) linked studies were identified by automated methods and 10 (8.1%) were identified only by a manual systematic review; 36 (29.0%) could not be linked to a published study.\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-6710551/v1/502aea376751f2b19613bbac.png"},{"id":83666194,"identity":"85328636-fbc5-464d-ab07-79b8a05aee86","added_by":"auto","created_at":"2025-05-30 11:41:46","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":47921,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMaximum-likelihood phylogenetic tree of 100 near-complete CCHF virus Large (L) segment sequences encompassing 10500 nucleotides\u003c/strong\u003e. Sequences were selected using the \u003cem\u003epplacer\u003c/em\u003e program to represent the known diversity of published L segment sequences (8). The tree was inferred using IQ-TREE 2 under \u0026nbsp;maximum likelihood criteria (9). Sequences marked with an asterisk were previously assigned to one of the six lineages indicated in the figure, in one or more published studies (10–13).\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-6710551/v1/8f051616fc020b15b5c03149.png"},{"id":83666197,"identity":"31632af7-df45-40ca-bb1b-557014cf5011","added_by":"auto","created_at":"2025-05-30 11:41:46","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":65139,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMaximum-likelihood phylogenetic tree of 100 near-complete Lassa virus RdRp (L) gene sequences encompassing 4600 nucleotides\u003c/strong\u003e. Sequences were selected using the \u003cem\u003epplacer\u003c/em\u003e program to represent the known diversity of published L gene sequences (8). The tree was inferred using IQ-TREE 2 under maximum likelihood criteria (9). Sequences marked with an asterisk were previously assigned to one of the seven lineages indicated in the figure, in a previously published study (15).\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-6710551/v1/957cfecf83c75bf687a44c54.png"},{"id":83666201,"identity":"1856c8d0-e65e-4c20-9da4-73e24b1a8bff","added_by":"auto","created_at":"2025-05-30 11:41:52","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":53407,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMaximum-likelihood phylogenetic tree of 79 complete Nipah virus nucleocapsid (N) gene sequences encompassing 1596 nucleotides\u003c/strong\u003e. The tree was inferred using IQ-TREE 2 under maximum likelihood criteria (9). The NiV-B cluster predominantly contains viruses isolated from human outbreaks in Bangladesh between 2004 and 2023 and from bats in Bangladesh, Thailand, and Cambodia. The NiV-I cluster contains viruses isolated from humans and bats in India between 2019 and 2023. The NiV-M cluster contains viruses isolated from the earliest recognized outbreak in Malaysia in 1999 and from bats in Malaysia, Thailand, and Cambodia isolated between 2003 and 2013.\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-6710551/v1/1a26814315d2a0b6dd83161e.png"},{"id":83666199,"identity":"44296120-f031-485a-bc91-46a591de45df","added_by":"auto","created_at":"2025-05-30 11:41:46","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":311493,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGPT-4 accuracy at extracting data from published research studies\u003c/strong\u003e. For the 223 CCHF virus, 72 Lassa virus, and 37 Nipah virus studies, the overall accuracy of GPT-4o was 91.2%, 92.3%, and 89.6%, respectively. An additional, 2.6%, 1.8%, and 3.5% of answers for CCHF, Lassa, and Nipah virus were partially correct. GPT-4o errors were more common when answering questions about the viral hosts and specimen types than about sample years, countries, sequenced genes, and accession numbers.\u003c/p\u003e","description":"","filename":"image7.png","url":"https://assets-eu.researchsquare.com/files/rs-6710551/v1/d0413d6406cf0569eccb66df.png"},{"id":96650103,"identity":"16268b6f-6e0f-4827-8cf3-59a5e6b693db","added_by":"auto","created_at":"2025-11-24 16:07:43","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3764553,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6710551/v1/558900d9-5f4f-43eb-b3e5-c6e0abca58a0.pdf"},{"id":83666202,"identity":"701c8f9b-efd1-4ee8-b91e-bc67608e248c","added_by":"auto","created_at":"2025-05-30 11:41:54","extension":"zip","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":90798,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFiles.zip","url":"https://assets-eu.researchsquare.com/files/rs-6710551/v1/7fb6a9448449760d7d876677.zip"}],"financialInterests":"No competing interests reported.","formattedTitle":"GenBank2PubMed: Bridging Viral Genomic Data and the Scientific Literature with AI-Assisted Curation","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eThe National Center for Biotechnology Information (NCBI) GenBank, DNA Databank of Japan (DDBJ), and the European Nucleotide Archive (ENA) are the three major public nucleotide sequence databases comprising the International Nucleotide Sequence Database Collaboration (INSDC) (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). These three databases\u0026mdash;hereafter collectively referred to as GenBank\u0026mdash;share data daily to ensure that their contents remain synchronized. Viral sequences in these databases are typically annotated with information about the host species, the specimen types from which the viruses were isolated, and, for human-infecting viruses, demographic data about the infected individuals. For treatable pathogenic human viruses, data on antiviral treatments received by individuals from whom the viruses were isolated are sometimes, but rarely, included.\u003c/p\u003e \u003cp\u003eFor the past 25 years, we have maintained the Stanford HIV Drug Resistance Database (Stanford HIVDB), a curated database containing HIV sequences annotated with antiviral treatment histories of the individuals from whom the viruses were obtained (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). To support this effort, we developed manual and computational approaches to annotate sequence data from GenBank with data extracted from research publications. Our software pipeline aggregates individual GenBank entries into submission sets likely to be linked based on the similarity of their authors, titles, and submission dates, as well as the sequential nature of their accession numbers.\u003c/p\u003e \u003cp\u003eWe have also long been interested in the extent to which published studies reporting virus sequences submit those sequences and their associated metadata to GenBank (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e). More recently, we have explored the use of automated methods\u0026mdash;including large language models (LLMs)\u0026mdash;to integrate data from published studies with the sequence data in GenBank (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn this study, we formalized and generalized the software pipeline originally developed for the Stanford HIVDB so that it could be applied to other pathogenic human viruses. We also developed additional automated tools, including LLMs, to link GenBank submission sets with their corresponding published studies \u0026ndash; a non-trivial task, given that many GenBank entries lack publication references. The result is a software toolkit that we call \u003cem\u003eGenBank2PubMed\u003c/em\u003e capable of rapidly generating a database in which all available gene sequences of a virus are annotated by their host, country and year of isolation.\u003c/p\u003e \u003cp\u003eFor this study, we selected three high-mortality viruses with outbreak potential \u0026ndash; Crimean Congo Hemorrhagic Fever Virus, Lass Virus, and Nipah Virus (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e). In parallel, we conducted a systematic review of the literature to assess the completeness of our automated methods at identifying published studies that were not included in the GenBank records. This manual review allowed us to determine how often sequences from published studies were submitted to GenBank.\u003c/p\u003e"},{"header":"METHODS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eDictionary of essential virus data\u003c/h2\u003e \u003cp\u003eFor each virus we created a dictionary of essential virus data including (i) the formal species name for the virus used by the NCBI; (ii) a list of each virus\u0026rsquo;s gene names and their synonyms; (iii) the NCBI virus reference sequence and its GenBank accession number; (iv) the formal names of virus host species mapped to commonly used names. For example, Homo sapiens was mapped to \u0026ldquo;humans\u0026rdquo;; various tick species were mapped to \u0026ldquo;ticks\u0026rdquo;; and bat species were mapped to \u0026ldquo;bats\u0026rdquo;.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eDefining GenBank submission sets\u003c/h3\u003e\n\u003cp\u003eWe searched GenBank using the following formal virus species names: Orthonairovirus haemorrhagiae for Crimean Congo Hemorrhagic Fever (CCHF) virus, Mammarenavirus lassaense for Lassa virus, and Henipavirus nipahense for Nipah virus. These searches yielded composite GenBank files containing the complete set of GenBank entries for each virus. The composite files were processed using the BioPython library to extract the GenBank reference, feature, descriptive, and sequence data for each entry. Individual GenBank entries with the same PubMed ID (PMID) or highly similar study titles, lists of authors, submission years, and sequential accession numbers were aggregated into GenBank submission sets, each containing one or more GenBank entries. Submission sets containing artificial constructs, patented sequences, unverified sequences, or laboratory manipulated viruses were identified by their GenBank annotation and excluded from further analysis.\u003c/p\u003e\n\u003ch3\u003eLinking GenBank submission sets to published studies\u003c/h3\u003e\n\u003cp\u003eGenBank submission sets containing a PMID were directly linked to the corresponding published study. For submission sets without a PMID, we employed two automated methods to identify the corresponding published study: (i) querying the PubMed API for publications authored by all pairs of contributors of the GenBank submission set and (ii) using the GPT-4o API to search the web for potential publications associated with the submission (Supplementary Text File). We then assessed whether these methods converged on the same published study. To confirm linkage, we retrieved the publication\u0026rsquo;s PDF and checked for matching GenBank accession numbers. If no accession numbers were present, we evaluated the study\u0026rsquo;s title, abstract, and author list to determine the likelihood of its association with the GenBank submission set. The software pipeline for aggregating GenBank entries into submission sets and to integrate the submission sets with the appropriate published studies is summarized in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eConfirmatory PubMed search and literature review\u003c/h3\u003e\n\u003cp\u003eTo evaluate the performance of our automated methods in linking published studies to GenBank submission sets, we conducted a manual PubMed search and literature review for each virus. We used the search terms \u0026ldquo;Crimean-Congo Hemorrhagic Fever Virus\u0026rdquo;, Lassa Virus\u0026rdquo;, and \u0026ldquo;Nipah Virus\u0026rdquo;. Two reviewers screened all titles and abstracts to identify publications potentially reporting viral sequence data. Flagged studies underwent a full-text review to confirm whether it indeed reported viral sequences. The manual PubMed search served an additional purpose. It enabled us to determine what proportion of published studies reporting viral sequences submitted the sequences to GenBank.\u003c/p\u003e \u003cp\u003eGenBank submission sets that could not be linked to a published study using either of our two automated methods or manual PubMed review were further investigated to determine if they had been published in a non-PubMed-indexed journal. For these cases, we searched SCOPUS using the virus name along with one or more of the submission set authors. We then reviewed the resulting publications to assess whether they referenced the same accession numbers or if the similarity in author lists and publication titles indicated that they were associated with the GenBank submission set.\u003c/p\u003e\n\u003ch3\u003eCreating a relational database\u003c/h3\u003e\n\u003cp\u003eWe created SQLite relational databases for each virus, each containing four core tables (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e): (i) tblGBEntries, which contains GenBank accession numbers and metadata associated with each GenBank entry; (ii) tblSequences, which contains separate rows for each gene in a GenBank entry and that entry\u0026rsquo;s nucleotide and amino acid sequences. The start and stop position of the amino acid sequence following alignment to the virus\u0026rsquo;s reference sequence for that gene and it\u0026rsquo;s percent identity to the reference sequence are included; (iii) tblGBSubmissionSetRefs, which contains reference data for each GenBank submission set; and (iv) tblPublications, which contains reference data for each published study.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe database also contains six supporting tables (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e): (i) tblGBRefLink, which manages the many-to-many relationship between tblGBSubmissionSetRefs and tblGBEntries; (ii) tblGBPubRefLink, which manages the occasional many-to-many relationship between tblGBSubmissionSets and tblPublications; (iii) tblPublicationsData, which stores the raw data extracted from publications; (iv) tblInsertions, which stores the insertions within each sequence relative to the virus\u0026rsquo;s reference genes; (v) tblQA, which stores the number of unsequenced bases, number of frameshifts, number of indels, number of stop codons, and number of differences from the reference sequence; and (vi) tblIsolates, which links virus isolates having more than one accession number, a situation that arises when authors create separate GenBank submissions for different genes of the same virus.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003ePhylogenetic analyses\u003c/h2\u003e \u003cp\u003eTo evaluate the consistency of our database with published phylogenetic analyses, we queried the database for near-complete gene sequences along with associated metadata including host, country, and sample year, and performed our own phylogenetic analyses. Each sequence was aligned to the reference nucleic acid sequence using Scikit-Bio (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://scikit.bio/\u003c/span\u003e\u003cspan address=\"https://scikit.bio/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). Insertions were excluded from the alignment, while deletions and ambiguous characters were retained. To limit the CCHF and Lassa virus trees to 100 sequences that represent the diversity of these virus populations, we used the pplacer program to select those sequences that minimize the average distance to the closest leaf (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e). Phylogenetic trees were constructed using IQTree2 with a general time reversible\u0026thinsp;+\u0026thinsp;gamma distribution model (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eData extraction from published studies\u003c/h3\u003e\n\u003cp\u003eFor each English-language published study determined to be reporting viral sequences, we extracted seven key data elements: (i) the host species from which samples were obtained; (ii) the type of specimen sequenced; (iii) the country or countries of the sequenced samples; (iv) the years in which virus samples were collected; (v) the viral genes sequenced; (vi) the method of sequencing; and (vii) the reported GenBank accession numbers. Data extraction was performed both manually and using GPT-4o. GPT-4o\u0026rsquo;s accuracy was defined as the frequency with which its extracted data matched the manual review. GPT-4o\u0026rsquo;s data extractions that did not exactly match the manual review were classified as either partially or completely incorrect. For example, a result was considered partially incorrect if GPT-4o identified the correct country but omitted another relevant country. However, if GPT-4o identified an incorrect country or misidentified the country entirely, the result was classified as completely incorrect.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eCCHF virus\u003c/h2\u003e \u003cp\u003e \u003cstrong\u003eGenBank submission sets and linkage to published studies\u003c/strong\u003e \u003cp\u003eAs of November 27, 2024, there were 4,848 GenBank entries. 94 entries were excluded because they represented artificial constructs, patented sequences, unverified sequences, or laboratory viruses. The remaining entries were aggregated into 193 submission sets containing 4,754 entries.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eOf the 193 submission sets, 173 (89.6%) were linked to a publication including 115 that contained a PMID (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA). Of the 78 submission sets without a PMID, 48 were linked to a published study through at least one of two automated methods: PubMed author query (n\u0026thinsp;=\u0026thinsp;27) and GPT-4o web search (n\u0026thinsp;=\u0026thinsp;39). An additional 9 submission sets were only identified by our manual PubMed search and literature review. One submission set that could not be linked to a published study by any other method was linked to a study by searching SCOPUS.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe 20 remaining submission sets that could not be linked to a published study contained 147 GenBank entries. Twelve had the title \u0026ldquo;Direct submission,\u0026rdquo; making it difficult to identify a corresponding publication. Four were submitted to GenBank between 2023 and 2024. Supplementary Table\u0026nbsp;1 provides details on the 78 CCHF virus GenBank submission sets without a linked PMID in GenBank and the methods by which they were linked to a published study.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRelational database\u003c/strong\u003e \u003cp\u003eOf the GenBank sequences, 3,417 (71.8%) were obtained from human samples, 990 (20.8%) from ticks, and 2.2% from livestock. For 5.1% of sequences, the host was not indicated in the GenBank record. All samples from humans in which the specimen type was provided were from blood, serum, or plasma.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eThe most common countries from which samples were obtained were Russia (46.8%), Turkey (14.3%), Pakistan (5.8%), India (4.7%), Kosovo (4.5%), Iran (3.7%), China (3.5%), South Africa (2.7%), Uganda (1.4%), Spain (1.1%), Kazakhstan (1.1%), and Bulgaria (1.0%). Thirty-five countries, each reporting less than 1% of samples, comprised 7.8% of GenBank sequences. For 1.6% of sequences no country could be identified. The sample year was available for 92.7% of sequences. The distribution of samples by decade was as follows: 1950\u0026ndash;1959 (0.2%), 1960\u0026ndash;1969 (1.3%), 1970\u0026ndash;1979 (1.2%), 1980\u0026ndash;1989 (2.9%), 1990\u0026ndash;1999 (0.7%), 2000\u0026ndash;2009 (17.6%), 2010\u0026ndash;2019 (69.2%), and 2020\u0026ndash;2024 (7.1%).\u003c/p\u003e \u003cp\u003eThe full genome comprising each of the three CCHF virus segments \u0026ndash; Large (L), Medium (M), and Small (S) \u0026ndash; was available for 4.9% of virus isolates. For 92.5% of isolates, a single segment was reported and for 2.6% of isolates, two segments were reported. Overall, the part of the S segment containing the nucleoprotein gene was reported for 52.4% of isolates while approximately 32.7% of isolates reported M segments and 27.3% reported the L segment.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003ePhylogenetic analyses\u003c/strong\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e shows the phylogenetic tree containing 100 near-complete L segments encompassing 10500 nucleotides. It demonstrates each of the six previously reported lineages (\u003cspan additionalcitationids=\"CR11 CR12\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). Although each lineage tends to occur in specific geographic regions, there are some exceptions postulated to arise from long distance virus transfers through migratory birds (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eLassa virus\u003c/h2\u003e \u003cp\u003e \u003cstrong\u003eGenBank submission sets and linkage to published studies\u003c/strong\u003e \u003cp\u003eAs of November 27, 2024, there were 2,780 GenBank entries. We excluded 117 entries because they represented artificial constructs, patented sequences, unverified sequences, or laboratory isolates. The remaining entries were aggregated into 78 submission sets containing 2,663 entries.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eOf the 78 submission sets, 64 (80.0%) were linked to a published study including 44 submission sets containing a PMID (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB). Of the 34 submission sets that did not contain a PMID, 21 were linked to a published study through at least one of two automated methods: PubMed author query (n\u0026thinsp;=\u0026thinsp;18) and GPT-4o web search (n\u0026thinsp;=\u0026thinsp;20). No additional submission set was identified by our manual searches of PubMed and SCOPUS.\u003c/p\u003e \u003cp\u003eThe 13 submission sets that could not be linked to a publication contained 171 entries. Six sets had the title \u0026ldquo;Direct submission\u0026rdquo; making it difficult to identify a corresponding published study. Two sets were submitted to GenBank between 2023 and 2024. Supplementary Table\u0026nbsp;2 provides details on the 34 Lassa virus GenBank submission sets without a linked PMID and the methods by which they were linked to a published study.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRelational database\u003c/strong\u003e \u003cp\u003eOf the 2,663 sequences in GenBank, 56.5% were from human samples, 37.9% were from rodents, and 0.3% were from other animals. For 5.3% the host was not known. For the 921 human samples for which the specimen type was provided, 98.8% were from blood, 0.7% from CSF, 0.2% from urine, 0.2% from pleural fluid, and 0.1% from breast milk.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eThe countries from which samples were obtained were Nigeria (54.3%), Guinea (18.1%), Sierra Leone (15.1%), Liberia (3.1%), Cote d'Ivoire (2.2%), Benin (1.2%), and Mali (1.0%). For 4.7% of samples, no country was identified. The sample year was available for 94.0% of sequences. The distribution of samples by decade was as follows: 1960\u0026ndash;1969 (0.1%), 1970\u0026ndash;1979 (0.7%), 1980\u0026ndash;1989 (0.8%), 1990\u0026ndash;1999 (0.1%), 2000\u0026ndash;2009 (16.7%), 2010\u0026ndash;2019 (78.9%), and 2020\u0026ndash;2024 (2.7%).\u003c/p\u003e \u003cp\u003eThe full genome of all four of the virus\u0026rsquo; proteins were available for 11.0% of isolates. For 50.8% of isolates a single gene sequence was available and for 38.2% of isolates, two or three gene sequences were available. Nucleocapsid gene sequences were available for 55.7% of sequences; G gene sequences for 55.1% of isolates; L gene sequences for 33.6% of isolates; and Z gene sequences for 27.3% of isolates.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003ePhylogenetic analyses\u003c/strong\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e shows the phylogenetic tree containing 100 near-complete L genes encompassing 4600 nucleotides. It reveals four distinct clusters from Nigeria, which have been designated lineages I to IV, a cluster from Mali designated lineage V, and several smaller clusters from Liberia, Sierra Leone, and Guinea, all classified under lineage IV (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e). Additionally, a cluster from Togo and Benin is identified as lineage VII (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e). Several of these lineages include isolates from both humans and rodents, with sequences clustering primarily by country.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eNipah virus\u003c/h2\u003e \u003cp\u003e \u003cstrong\u003eGenBank submission sets and linkage to published studies\u003c/strong\u003e \u003cp\u003eAs of November 27, 2024, there were 407 GenBank entries. 52 entries were excluded because they represented artificial constructs, patented sequences, unverified sequences, or laboratory viruses. The remaining entries were aggregated into 34 submission sets containing 355 entries.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eOf the 34 submission sets, 31 (91.2%) were linked to a publication including 22 submission sets containing a PMID (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC). Of the 12 submission sets without a PMID, 9 were linked to a published study through at least one of two automated methods: PubMed author query (n\u0026thinsp;=\u0026thinsp;4) and GPT-4o web search (n\u0026thinsp;=\u0026thinsp;9). No additional submission set was identified by our manual searches of PubMed and SCOPUS.\u003c/p\u003e \u003cp\u003eThree submission sets containing 33 GenBank entries could not be linked to a published study. One had the title \u0026ldquo;Direct submission\u0026rdquo; making it difficult to identify a corresponding published study. One set was submitted in 2024. Supplementary Table\u0026nbsp;3 provides details on the 12 Nipah virus GenBank submission sets without a linked PMID and the methods by which they were linked to a published study.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRelational database\u003c/strong\u003e \u003cp\u003eAmong the 355 sequences in GenBank, 62.5% were from bats, 32.7% were from humans, 1.4% from pigs, and 0.3% from dogs. For 3.1% (n\u0026thinsp;=\u0026thinsp;11) the host could not be determined. For the 134 human samples for which the specimen type was provided, 48.6% were from throat swabs or saliva, 45.5% were from urine, 3.7% were from blood, 1.5% were from CSF, and 0.7% were from breast milk.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eThe countries from which human and bat samples were obtained were from Bangladesh (36.1%), Thailand (29.0%), India (17.7%), Cambodia (9.0%), Malaysia (4.5%), Indonesia (1.4%), and Sri Lanka (1.1%). For 1.1% of sequences the country could not be determined. The sample year was available for 91.3% of GenBank submissions: 1998\u0026ndash;1999 (1.9%), 2000\u0026ndash;2009 (28.1%), 2010\u0026ndash;2019 (57.1%), 2020\u0026ndash;2024 (13.0%).\u003c/p\u003e \u003cp\u003eThe full genome containing each of the six Nipah virus proteins was available for 31.5% of isolates. For 61.4% of isolates, just the nucleocapsid gene was sequenced. For the remaining 7.1% of sequences one or more genes other than nucleocapsid was sequenced including G, M, P, L, or F.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003ePhylogenetic analyses\u003c/strong\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e shows the phylogenetic tree containing 79 complete N genes encompassing 1596 nucleotides. It demonstrates that sequences from Bangladesh and India comprise one major lineage while sequences from Malaysia form a second major lineage. The lineage containing sequences from Bangladesh and India contains two sublineages from Bangladesh and one sublineage from India. Sequences belonging to different hosts (i.e., human and bat) cluster by country rather than host. Sequences from bats in Thailand and Cambodia clustered with the Malaysian sequences.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eConfirmatory PubMed search and literature review\u003c/h2\u003e \u003cp\u003e \u003cstrong\u003eCCHF virus\u003c/strong\u003e \u003cp\u003eA PubMed search using the term \u0026ldquo;Crimean Congo Hemorrhagic Fever virus\u0026rdquo; performed December 10, 2025, retrieved 2,131 records. After reviewing titles and abstracts, 261 publications underwent full-text review and 181 were confirmed to be reporting CCHF virus sequences. Among these, 145 (80.1%) were successfully matched to a GenBank submission set. As noted above, literature searches identified ten additional matching studies that could not be linked through automated methods alone (nine via the PubMed search and one via the SCOPUS search).\u003c/p\u003e \u003c/p\u003e \u003cp\u003eOf the remaining 36 studies (19.9%) that could not be matched to a GenBank submission set, two contained sequences that were submitted to the NCBI Short Read Archive (SRA), two reported submitting their sequences but did not provide accession numbers, and two reported that their sequences were available upon request. The 36 CCHF virus PubMed publications without GenBank-linked sequences is provided in Supplementary Table\u0026nbsp;4.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eLassa virus\u003c/strong\u003e \u003cp\u003eA PubMed search using the term \u0026ldquo;Lassa virus\u0026rdquo; performed January 6, 2025, retrieved 1,680 records. After reviewing titles and abstracts, 103 publications underwent full-text review and 67 were confirmed to be reporting Lassa virus sequences. Among these, 58 (86.6%) were successfully matched to a GenBank submission set. The PubMed search did identified one additional study that could not be linked through automated methods alone.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eOf the remaining nine studies (13.4%) that could not be matched to a GenBank submission set, two were case reports. The nine Lassa virus PubMed publications without GenBank-linked sequences is provided in Supplementary Table\u0026nbsp;5.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eNipah virus\u003c/strong\u003e \u003cp\u003eA PubMed search using the term \u0026ldquo;Nipah virus\u0026rdquo; performed December 15, 2024, retrieved 1,515 records. After reviewing titles and abstracts, 76 publications underwent full-text review and 32 were confirmed to be reporting Nipah virus sequences. Among these, 28 (87.5%) were successfully matched to a GenBank submission set. The PubMed search did not identify additional studies that could not be linked through automated methods alone.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eOf the remaining four studies (12.5%) that could not be matched to a GenBank submission set, one reported three sequences and stated that the sequences had been submitted to GenBank but did not provide accession numbers. The four Nipah virus PubMed publications without GenBank-linked sequences are listed in Supplementary Table\u0026nbsp;6.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eAccuracy of GPT-4 data extraction\u003c/h2\u003e \u003cp\u003eFor each of the published studies, we assessed the accuracy of GPT-4 at extracting information on the hosts, specimen types, sequencing method, sample years, countries, genes, and GenBank accession numbers. Figure\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e shows GPT-4 accuracy for each data type for each virus. For the 223 CCHF virus, 72 Lassa virus, and 37 Nipah virus studies, the overall accuracy of GPT-4 was 91.2%, 92.3%, and 89.6%, respectively. An additional 2.6%, 1.8%, and 3.5% of questions for each virus were partially correct.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe data extracted from published studies modestly reduced the proportion of GenBank entries for which the host, country, or sample year was unknown. The proportion of entries for which the host species, country, or sample year were missing for the three viruses combined based on the GenBank record alone was 4.3%, 5.3%, and 4.3%, respectively. Following extraction of data from published papers, these proportions were 0.6%, 0.7%, and 0.1%, respectively.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eResources\u003c/h2\u003e \u003cp\u003eWe created a dedicated website for this project, which provides comprehensive information for each virus: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://hivdb.stanford.edu/genbank2pubmed/\u003c/span\u003e\u003cspan address=\"https://hivdb.stanford.edu/genbank2pubmed/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. The website includes: (i) A link to download the virus\u0026rsquo;s SQLite database; (ii) A link to download Excel files summarizing each virus\u0026rsquo;s GenBank submission sets; (iii) A dynamic table summarizing metadata for each virus\u0026rsquo;s GenBank submission set, including virus host, country, specimen, year, and sequenced genes; (iv) A static list of GenBank submission sets linked to a published study; (v) A static list of those GenBank submission sets not linked to a published study; (vi) A static list of published studies reporting sequences not found in GenBank; (vii) Phylogenetic analyses of viral gene sequences.\u003c/p\u003e \u003c/div\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eWe developed a software pipeline called \u003cem\u003eGenBank2PubMed\u003c/em\u003e that created annotated relational databases for three high-mortality viruses with outbreak potential. The pipeline consists of the following steps. First, Python scripts group GenBank entries into submission sets based on highly similar titles, author lists, submission years, and sequential accession numbers. Only those GenBank entries that represent clinical or zoonotic viruses were included while patented sequences, viral constructs, and laboratory isolates were excluded. Second, two automated methods identify published studies associated with those submission sets for which the GenBank entries lack PMIDs. Third a manual PubMed search is performed to identify studies that were not identified by our automated methods and to obtain metadata that was not already contained in GenBank. Fourth, data from GenBank entries and linked published studies are integrated into a relational database.\u003c/p\u003e \u003cp\u003eTo create the GenBank submission sets, we processed GenBank files containing all virus entries, ranging in size from 4 mb for Nipah virus to 25 mb for CCHF virus, converting them into Excel files ranging in size from 17 kb to 68 kb that can be browsed by reference data, sequenced gene, host, country, and year.\u003c/p\u003e \u003cp\u003eAcross all three viruses, we identified a total of 305 GenBank submission sets. Of these, 181 (59.3%) included a PMID in their GenBank entries, while 124 (40.7%) lacked any link to a published study. For these 124 sets, automated PubMed searches and GPT-4 web searches combined to identify 78 (62.9%) linked published studies. Manual literature searches yielded just 10 (8.1%) linked studies that were not found by automated methods. The remaining 36 submission sets – 29.0% of those lacking PMIDs and 11.8% of the total – could not be associated with a publication; among these were 19 labeled “Direct submission” and seven submitted to GenBank between 2023 and 2024.\u003c/p\u003e \u003cp\u003eThe relational databases created are useful to researchers determining the extent of virus sequencing by host, country, sample year, and gene. This may influence the design of surveillance studies to address gaps in publicly available virus sequence data. It also facilitates the selection of sequences for phylogenetic analyses. Because the sequences are stored in a virtual alignment to established reference sequences for each virus gene, it is also possible to identify those sequences encompassing sufficient coverage for phylogenetic analysis. The database also contains quality control data for each sequence such as the number of indels, stop codons, unsequenced bases, and frameshifts which can be used to exclude certain sequences from analysis.\u003c/p\u003e \u003cp\u003eAlthough the manual PubMed search identified only ten additional publications not captured by automated methods, it allowed us to address a related question: how frequently do authors who report viral sequences in their publications also submit those sequences to GenBank? We found that 80.1% of CCHF virus studies, 86.6% of Lassa virus studies, and 87.5% of Nipah virus studies that reported sequences had corresponding submissions in GenBank. This high rate of GenBank submission likely reflects the relatively limited number of sequences available for these viruses, in contrast to more extensively sequenced pathogens. For comparison, we previously reported that approximately 60% of 934 HIV drug resistance studies published between 2010 and 2019 had submitted sequence data to GenBank (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn recent years, there have been multiple studies investigating the application of LLMs to a range of tasks related to the processing of biomedical literature, including publication triage, content summarization, and data extraction (\u003cspan additionalcitationids=\"CR18\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e–\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e). In this study, we used GPT-4o to identify published studies linked to GenBank submission sets and to extract data from published studies. We found that its accuracy ranged from about 80–95% depending on the type of data extracted. We expect that this accuracy will improve over time as LLMs improve and potentially with fine-tuning, a process that entails the continued training of a pre-trained LLM on domain-specific material for the purpose of adapting the model to specific tasks (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e). Indeed, we have recently reported that the accuracy of GPT-4-mini and the open source model Llama 3.1 Instruct 70B at answering questions about studies on HIV drug resistance can be significantly improved by fine-tuning (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e)\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eLimitations\u003c/strong\u003e \u003c/p\u003e\u003cp\u003eFirst, because LLMs can be inaccurate, their findings required manual verification. Published studies linked to GenBank submission sets by automated methods had to be manually reviewed if they lacked matching accession numbers. Metadata extracted from publications also needed to be manually validated for accuracy. Second, 11.8% (36/305) of GenBank submission sets could not be linked to a published study using any approach, including two automated methods, a manual PubMed search, and a final SCOPUS search. These sequences may have been described in studies we were unable to identify. Finally, although identifying the corresponding studies provided useful context, the amount of additional epidemiological data extracted was smaller than expected; the proportion of sequences lacking such data declined only modestly – from approximately 5–1% – after data extraction.\u003c/p\u003e \u003cp\u003e\u003c/p\u003e \u003cp\u003e\u003c/p\u003e"},{"header":"Conclusions and future directions","content":"\u003cp\u003eThe creation of GenBank submission sets makes it possible for the large number of GenBank entries for a virus to be consolidated into a tractable number of submission sets containing just those viruses of epidemiological relevance. Automated methods including an LLM can identify published studies that describe the GenBank submission sets in more detail than indicated by the GenBank entries alone. Finally, as LLMs improve at performing systematic reviews – specifically identifying published papers and extracting specific data – the process of creating virus sequence databases that include both data from GenBank and from published studies will be streamlined.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCCHF\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCrimean-Congo Hemorrhagic Fever\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eNCBI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eNational Center for Biotechnology Information\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDDBJ\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDNA Databank of Japan\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eENA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eEuropean Nucleotide Archive\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eINSDC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eInternational Nucleotide Sequence Database Collaboration\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLLM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eLarge language model\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ePMID\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ePubMed ID\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSRA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eShort Read Archive\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":" \u003cp\u003e \u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e \u003cp\u003eNot applicable. This study only utilized publicly available data.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent for publication\u003c/strong\u003e \u003cp\u003eNot applicable. This study only utilized publicly available data.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThe research reported in this paper was supported by a grant from the NIH: NIH/NIAID R24AI13661806.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eKT and RWS conceived of the project; KT and JZ wrote the software for the project; KT, JZ, YM, and RWS performed literature reviews; KT, JZ, YM, and RWS performed data analysis; KT, JZ, YM, and RWS created one or more figures; RWS wrote the manuscript; All authors reviewed the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eNot applicable\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe data in this study are available online (https://hivdb.stanford.edu/genbank2pubmed/), in a GitHub repository (https://github.com/hivdb/GenBankRefs), and in six supplementary files.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eKarsch-Mizrachi, I., Takagi, T. \u0026amp; Cochrane, G. The international nucleotide sequence database collaboration. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cb\u003e46\u003c/b\u003e (Database issue), D48\u0026ndash;51 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShafer, R. W., Jung, D. R., Betts, B. J., Xi, Y. \u0026amp; Gonzales, M. J. Human Immunodeficiency Virus Reverse Transcriptase and Protease Sequence Database. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cb\u003e28\u003c/b\u003e (1), 346\u0026ndash;348 (2000).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRhee, S. Y. et al. Public availability of HIV-1 drug resistance sequence and treatment data: a systematic review. \u003cem\u003eLancet Microbe\u003c/em\u003e. \u003cb\u003e3\u003c/b\u003e (5), e392\u0026ndash;e398 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eInzaule, S. C. et al. Recommendations on data sharing in HIV drug resistance research. \u003cem\u003ePLoS Med.\u003c/em\u003e \u003cb\u003e20\u003c/b\u003e (9), e1004293 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTao, K. et al. GPT-4 performance on querying scientific publications: reproducibility, accuracy, and impact of an instruction sheet. \u003cem\u003eBMC Med. Res. Methodol.\u003c/em\u003e \u003cb\u003e24\u003c/b\u003e (1), 139 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTao, K. et al. Fine-tuned large language models for answering questions about full-text biomedical research studies [Internet]. medRxiv; 2024 [cited 2025 Mar 23]. p. 2024.10.28.24316263. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.medrxiv.org/content/\u003c/span\u003e\u003cspan address=\"https://www.medrxiv.org/content/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1101/2024.10.28.24316263v2\u003c/span\u003e\u003cspan address=\"10.1101/2024.10.28.24316263v2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMehand, M. S., Al-Shorbaji, F., Millett, P. \u0026amp; Murgue, B. The WHO R\u0026amp;D Blueprint: 2018 review of emerging infectious diseases requiring urgent research and development efforts. \u003cem\u003eAntiviral Res.\u003c/em\u003e \u003cb\u003e159\u003c/b\u003e, 63\u0026ndash;67 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMatsen, F. A. I. V., Gallagher, A. \u0026amp; McCoy, C. O. Minimizing the Average Distance to a Closest Leaf in a Phylogenetic Tree. \u003cem\u003eSyst. Biol.\u003c/em\u003e \u003cb\u003e62\u003c/b\u003e (6), 824\u0026ndash;836 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMinh, B. Q. et al. IQ-TREE 2: New Models and Efficient Methods for Phylogenetic Inference in the Genomic Era. \u003cem\u003eMol. Biol. Evol.\u003c/em\u003e \u003cb\u003e37\u003c/b\u003e (5), 1530\u0026ndash;1534 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCarroll, S. A., Bird, B. H., Rollin, P. E. \u0026amp; Nichol, S. T. Ancient common ancestry of Crimean-Congo hemorrhagic fever virus. \u003cem\u003eMol. Phylogenet. Evol.\u003c/em\u003e \u003cb\u003e55\u003c/b\u003e (3), 1103\u0026ndash;1110 (2010).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBente, D. A. et al. Crimean-Congo hemorrhagic fever: History, epidemiology, pathogenesis, clinical syndrome and genetic diversity. \u003cem\u003eAntiviral Res.\u003c/em\u003e \u003cb\u003e100\u003c/b\u003e (1), 159\u0026ndash;189 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLukashev, A. N. et al. Phylogeography of Crimean Congo Hemorrhagic Fever Virus. \u003cem\u003ePLOS ONE\u003c/em\u003e. \u003cb\u003e11\u003c/b\u003e (11), e0166744 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShahhosseini, N. et al. Crimean-Congo Hemorrhagic Fever Virus in Asia, Africa and Europe. \u003cem\u003eMicroorganisms\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e (9), 1907 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLindeborg, M. et al. Migratory Birds, Ticks, and Crimean-Congo Hemorrhagic Fever Virus - Volume 18, Number 12\u0026mdash;December 2012 - Emerging Infectious Diseases journal - CDC. [cited 2025 Apr 2]; Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://wwwnc.cdc.gov/eid/article/18/12/12-0718_article\u003c/span\u003e\u003cspan address=\"https://wwwnc.cdc.gov/eid/article/18/12/12-0718_article\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eForni, D. \u0026amp; Sironi, M. Population structure of Lassa Mammarenavirus in West Africa. \u003cem\u003eViruses\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e (4), 437 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGarry, R. F. Lassa fever \u0026mdash; the road ahead. \u003cem\u003eNat. Rev. Microbiol.\u003c/em\u003e \u003cb\u003e21\u003c/b\u003e (2), 87\u0026ndash;96 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDennst\u0026auml;dt, F., Zink, J., Putora, P. M., Hastings, J. \u0026amp; Cihoric, N. Title and abstract screening for literature reviews using large language models: an exploratory study in the biomedical domain. \u003cem\u003eSyst. Rev.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e (1), 158 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLieberum, J. L. et al. Large language models for conducting systematic reviews: on the rise, but not yet ready for use\u0026mdash;a scoping review. Journal of Clinical Epidemiology [Internet]. 2025 May 1 [cited 2025 Mar 28];181. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.jclinepi.com/article/S0895-4356(25)00079-4/fulltext\u003c/span\u003e\u003cspan address=\"https://www.jclinepi.com/article/S0895-4356(25)00079-4/fulltext\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSantos, \u0026Aacute;. O., da Silva, E. S., Couto, L. M., Reis, G. V. L. \u0026amp; Belo, V. S. The use of artificial intelligence for automating or semi-automating biomedical literature analyses: A scoping review. \u003cem\u003eJ. Biomed. Inform.\u003c/em\u003e \u003cb\u003e142\u003c/b\u003e, 104389 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu, E. J. et al. LoRA: Low-Rank Adaptation of Large Language Models [Internet]. arXiv; [cited 2024 Jul 5]. (2021). Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://arxiv.org/abs/2106.09685\u003c/span\u003e\u003cspan address=\"http://arxiv.org/abs/2106.09685\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTinn, R. et al. Fine-tuning large neural language models for biomedical natural language processing. \u003cem\u003ePatterns\u003c/em\u003e \u003cb\u003e4\u003c/b\u003e (4), 100729 (2023).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Virus sequences, molecular epidemiology, phylogenetics, GenBank, PubMed, GPT-4o, Crimean Congo Hemorrhagic Fever virus, Lassa virus, Nipah virus","lastPublishedDoi":"10.21203/rs.3.rs-6710551/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6710551/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cu\u003e\u003cstrong\u003eBackground\u003c/strong\u003e\u003c/u\u003e: GenBank entries of pathogenetic viral sequences are typically annotated with host species and epidemiological metadata. However, linking these entries to their corresponding published studies remains labor-intensive.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cu\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/u\u003e: We developed \u003cem\u003eGenBank2PubMed\u003c/em\u003e, a computation pipeline that integrates GenBank sequence data with metadata from published studies. The pipeline aggregates GenBank entries into submission sets based on shared authorship, title similarity, submission dates, and the sequential nature of their accession numbers. Using automated methods, including GPT-4, we linked these submission sets to relevant publications – a challenging task given that many GenBank entries lack citation references. The result is a database in which viral sequences are annotated by host, country, and year of isolation. We also conducted a systematic review to assess how frequently published studies reporting sequences included GenBank submissions. We applied \u003cem\u003eGenBank2PubMed \u003c/em\u003eto three high-mortality viruses with outbreak potential: Crimean-Congo Hemorrhagic Fever (CCHF) virus, Lassa virus, and Nipah virus.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cu\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/u\u003e: We identified 193 CCHF virus submission sets (4,754 entries), 78 Lassa virus sets (2,663 entries), and 34 Nipah virus sets (355 entries). Of these, 173 (CCHF), 64 (Lassa), and 31 (Nipah) were linked to published studies. Integration with publication data enriched the contextual and epidemiological metadata for each set. Additionally, our literature review found that 80.1% of CCHF, 86.6% of Lassa, and 87.5% of Nipah virus studies reporting sequences had corresponding GenBank submissions. GenBank submission sets and relational databases for each virus are available at https://hivdb.stanford.edu/genbank2pubmed/; the pipeline is available at https://github.com/hivdb/GenBankRefs.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cu\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e\u003c/u\u003e: Creating submission sets facilitates the organization of GenBank data into browsable spreadsheets and queryable databases. GPT-4 contributed to linking GenBank entries with published studies and extracting metadata, although manual validation remained essential for accuracy. \u003cem\u003eGenBank2PubMed \u003c/em\u003erepresents a significant step toward integrating GenBank viral sequences with the scientific literature in which they are reported.\u003c/p\u003e","manuscriptTitle":"GenBank2PubMed: Bridging Viral Genomic Data and the Scientific Literature with AI-Assisted Curation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-30 11:41:41","doi":"10.21203/rs.3.rs-6710551/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-06-23T18:06:55+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-17T01:56:14+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-11T16:22:54+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"278899160014664438977290289619187742415","date":"2025-06-04T04:30:13+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"232754910136376059164761361070157378499","date":"2025-06-03T16:06:07+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"3821932723075546627632617583820429654","date":"2025-05-28T12:39:49+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-28T12:15:56+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-05-27T14:22:52+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-05-22T05:45:06+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-05-21T12:59:00+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-05-20T19:27:10+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f14a4a6b-6d2b-43ae-a462-d62eb7fa4398","owner":[],"postedDate":"May 30th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":49250836,"name":"Health sciences/Diseases/Infectious diseases/Viral infection"},{"id":49250837,"name":"Biological sciences/Computational biology and bioinformatics/Databases/Genetic databases"},{"id":49250838,"name":"Biological sciences/Computational biology and bioinformatics/Data acquisition"},{"id":49250839,"name":"Biological sciences/Computational biology and bioinformatics/Data integration"},{"id":49250840,"name":"Biological sciences/Computational biology and bioinformatics/Literature mining"},{"id":49250841,"name":"Biological sciences/Computational biology and bioinformatics/Sequence annotation"}],"tags":[],"updatedAt":"2025-11-24T16:01:16+00:00","versionOfRecord":{"articleIdentity":"rs-6710551","link":"https://doi.org/10.1038/s41598-025-28386-8","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2025-11-22 15:57:09","publishedOnDateReadable":"November 22nd, 2025"},"versionCreatedAt":"2025-05-30 11:41:41","video":"","vorDoi":"10.1038/s41598-025-28386-8","vorDoiUrl":"https://doi.org/10.1038/s41598-025-28386-8","workflowStages":[]},"version":"v1","identity":"rs-6710551","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6710551","identity":"rs-6710551","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}