Public Omics Explorer (POE): Создание возможностей для комплексного семантического поиска по наборам данных GEO omics на основе публикаций PubMed.

Introduction

In the era of multi-omics research, repositories such as the NCBI Gene Expression Omnibus (GEO) [1] aggregate massive volumes of data. As of September 2025, GEO hosts more than 9.6 million samples and is expanding exponentially. However, despite this wealth, discovering datasets relevant to a specific biological query remains challenging. The core obstacle is the heterogeneity and often unstructured nature of metadata, which are typically expressed as free text using inconsistent vocabularies, despite the existence of well-documented submission guidelines since 2002 [2] and the establishment of the FAIR principles (Findable, Accessible, Interoperable, Reusable) [3] .

Early efforts to improve metadata consistency focused on manual curation. Notable examples include the BioSamples [4] and METAGENOTE [5] databases, as well as tools like ArcheGEO [6] GEOmetacuration [7] and Zooma ( https://www.ebi.ac.uk/spot/zooma/ ). The CEDAR Workbench [8] further supports the creation of semantically structured metadata templates based on biomedical ontologies [9] . While these initiatives have improved metadata quality and searchability, their reliance on human effort limits scalability and long-term sustainability.

To address these constraints, automated approaches have been developed using heuristics, rule-based logic, and machine learning. Tools such as metaSRA [10] , ALE [11] and ReGEO [12] utilize machine learning techniques to infer ontology terms or extract semantic structure. Others like tidyGEO [13] employ rule-based or interactive transformation methods to convert free-text metadata into structured, ontology-aligned fields. Complementary platforms such as GeoQuery [14] , GEOmetadb [15] , SRA-Explorer ( https://sra-explorer.info/ ), and SpiderSeqR [16] facilitate programmatic or SQL-like access to GEO, improving access to curated data subsets.

In parallel, an alternative line of tools focuses not on metadata but on expression-level similarity. Signature-based search tools, including LCE (Lung Cancer Explorer) [17] Expression Atlas [18] , LINCS [19] , and ARCHS4 [20] , support dataset retrieval based on gene expression profiles. Broader platforms like OmicsDI [21] and DisGeNET [22] support retrieval of datasets based on phenotype similarity, genetic variants, or interaction networks.

Advances in semantic search have enabled more meaningful retrieval, moving beyond keyword-based approaches to techniques that consider user intent and contextual understanding. Tools like PubTator3 [23] , LitCovid + GEO [1] , and LitSense [24] facilitate literature-level search using semantic embeddings or concept-based representations. In addition, BioASQ [25] offers a large-scale biomedical question-answering and information retrieval benchmark, focusing on semantic matching between PubMed abstracts and ontology-derived concepts. While these tools have advanced semantic exploration of biomedical text, they operate exclusively at the publication level, without linking retrieved articles to underlying experimental datasets. Related infrastructures, including EBI BioStudies [26] , Semantic Scholar ( https://www.semanticscholar.org/ ,) and TogoGenome [27] , enrich dataset and entity records with semantic annotations linked to biomedical ontologies such as Medical Subject Headings (MeSH) and Unified Medical Language System (UMLS).

Despite the often explicit linkage between GEO datasets and associated publications in PubMed or PMC, few tools have systematically used the content of these articles as a primary source of semantic information for dataset discovery. Biochat [28] , a prototype released as a preprint on bioRxiv, attempted to calculate similarity between GEO datasets using vector representations of associated article abstracts. However, it was never peer-reviewed and the service is now defunct. Similarly, LitCovid + GEO relied on keyword matching and manual curation, while PubTator3 and LitSense focused on extracting and indexing biomedical entities from literature without integrating dataset retrieval capabilities. Even official PubMed-GEO linkouts remain underutilized in semantic or conceptual search frameworks.

To address these limitations, we developed the Public Omics Explorer (POE) - a web application that enables semantic search across GEO datasets based on PubMed publications. POE systematically processes the titles and abstracts of articles linked to GEO datasets as input for publication‑content‑aware semantic search. It constructs a unified embedding space using SBioBERT, a biomedical-specialized transformer model pretrained on large-scale biomedical corpora to effectively capture domain-specific terminology and context. The model generates dense vector representations of each article’s content, which are indexed using Facebook AI Similarity Search (FAISS) to enable high-precision, context-aware retrieval of semantically related literature and associated datasets.

In addition to semantic search, POE offers structured filtering across multiple metadata dimensions, including organism, experiment type, library strategy, sample type, extracted molecule, and publication year. For users with prior knowledge of specific datasets or publications, the platform supports targeted retrieval via accession identifiers such as GSE IDs, PubMed IDs, or DOIs, providing direct links to the corresponding records hosted in external repositories like GEO and ENA [29] and their mirrored SRA records.

POE also includes a RESTful API and programmatic interface that allow integration into external pipelines, applications, or institutional infrastructures, supporting scalable and reproducible workflows in computational biology. Importantly, POE does not store or replicate omics datasets locally, but instead facilitates semantic access and redirection to their sources. The platform is updated daily through automated scheduling, ensuring that new GEO metadata ENA/SRA cross-references and PubMed records are continuously incorporated into the semantic index.

Methods

System overview and accessibility

POE is hosted as an online platform accessible through modern web browsers without the need for installation or user-side configuration ( Fig. 1 ). All computational infrastructure, including metadata ingestion, semantic embedding, indexing, and result delivery, is maintained on the server side.

Data sources and acquisition

POE integrates data from three major biomedical resources: GEO, PubMed, and the European Nucleotide Archive (ENA) [29] . GEO metadata is retrieved daily through NCBI’s FTP server and E-utilities interface. During each cycle, POE downloads and parses metadata for both study-level records (GSE) and sample-level records (GSM), extracting key fields such as organism, library strategy, experiment type, sample type, and molecule extracted. This metadata forms the core of the system's searchable omics dataset repository ( Fig. 2 ).

PubMed articles linked to GEO datasets are also collected during this process. When GSE records contain associated PubMed identifiers (PMIDs), the corresponding publications are programmatically retrieved. Metadata is extracted using a hybrid strategy that includes both NCBI API access and HTML parsing with the BeautifulSoup library [30] . For each article, POE collects the title, abstract, MeSH terms, author list, affiliations, publication dates, reported substances, and a list of similar articles. This comprehensive metadata profile is used to semantically index the literature and establish links to relevant omics datasets. Additionally, POE integrates metadata from the ENA, particularly to retrieve FASTQ file links associated with GEO datasets.

Data processing and integration pipeline

The integration of omics datasets with literature information in POE is managed by a custom-developed, fully automated pipeline executed daily. This pipeline updates the GEO and PubMed metadata stores, links GSE entries to associated publications, generates vector embeddings for newly indexed abstracts, and constructs a FAISS index to support semantic search.

For each new or updated GSE entry, the system attempts to map a PubMed article using the provided PMID. If a valid link is found, the corresponding publication metadata is fetched and associated with the dataset. Entries without a valid PubMed link are retained in the GEO index but excluded from semantic embedding. All structured metadata is normalized and stored in a PostgreSQL database. The schema is designed to be flexible and extensible, enabling advanced filtering across heterogeneous metadata fields and seamless integration with the semantic search backend.

Comparative evaluation of POE and standard GEO keyword search

Following the description of POE’s architecture and data integration pipeline, we conducted a systematic evaluation to quantify its performance relative to standard keyword-based GEO searches. The goal of this evaluation was to assess the added value of semantic search for dataset discovery. To ensure comparability, the analysis was restricted to PubMed-linked GEO datasets, which represent the subset directly supported by POE’s indexing strategy. We compared POE’s semantic search results with those retrieved from the GEO Datasets and GEO Profiles databases using an identical set of 30 natural-language biological queries. (see Supplementary Table 1 ).

These queries were organized into five thematic categories: a) Epigenetics, b)Immune regulation and cancer, c)Neurological disease, d)Non-coding RNA biology and e) Regulatory genomics and stem-cell development -to capture a broad spectrum of biological contexts.

Representative examples include “chromatin accessibility in stem-cell fate decisions” (Epigenetics), “neuroinflammation signatures in Alzheimer’s disease” (Neurological disease), and “miRNA promoter variants in cancer” (Non-coding RNA biology).

Each query was formulated manually by two domain experts to reflect realistic biological research questions that scientists might pose when exploring public omics data.

Queries were expressed in natural-language form, typically as concise phrases of 4–10 words, without Boolean operators or database-specific syntax.

Topics were selected to represent a balance between well-studied and emerging biological areas, covering diseases, molecular mechanisms, and regulatory processes.

This formulation ensured that the benchmark reflected authentic user behavior and evaluated POE’s strength in handling open-ended, concept-level queries rather than keyword matching.

This diversity ensured that the evaluation reflected dataset discovery across both established and emerging areas of biomedical research.

Quantitative analysis

For each biological query, identical searches were executed in both systems. Detailed results of each system can be found in Supplementary material 1 . The GEO search results were subsequently partitioned into two categories: (i) datasets associated with a PubMed identifier and (ii) datasets lacking a PubMed identifier. Since POE exclusively indexes datasets linked to PubMed publications, the comparative analysis was restricted to the PubMed-linked subset of GEO results.

Formally, for each query i let:

denote the set of PubMed-linked GEO datasets, and

denote the corresponding POE results.

Since POE typically yields a larger number of results than GEO ( ∣P i ∣>∣G i ∣) , we applied a truncation procedure to ensure equal-sized sets for pairwise comparison. Specifically, we defined

where top ∣Gi∣ (P i ) corresponds to the ∣G i ∣ elements of P i , ranked according to their semantic similarity scores to the query.

All subsequent analyses were therefore performed on the aligned pairs (G i ,P i ∗ ) . To further characterize the relative contributions of each system, we defined:

These categories respectively capture datasets identified only by GEO, only by POE, and by both systems for query i. The relative sizes of these sets were used to quantify overlap and divergence between the two search strategies. (see Supplementary Table 2 ).

Qualitative relevance assessment

To complement this overlap analysis, we performed a manual evaluation of the unique results from each system, that is, datasets belonging to GEO-unique or POE-unique. Each PubMed-linked dataset in these categories was scored for relevance by evaluating the associated publication title and abstract on a 1–10 scale. For statistical robustness, we restricted this evaluation to queries with more than four unique datasets, which resulted in the inclusion of 9 out of the 30 queries (see supplementary table 3 ).

Scoring followed a structured rubric comprising five criteria:

C1: Conceptual match, alignment between the query concept and the study focus (2 = direct, 1 = partial, 0 = keyword-only).

C2: Biological context, correspondence of species, tissue, disease, or biological system (2 = full match, 1 = partial, 0 = mismatch). C3: Methods/data fit, appropriateness of the experimental methods and data type to the query (2 = ideal, 1 = related, 0 = irrelevant).

C4: Contribution of findings, degree to which the study provides insights directly addressing the query (2 = direct, 1 = indirect/limited, 0 = none/theoretical).

C5: Focus/explicitness, whether the query topic was explicitly targeted versus only peripheral within a broader theme (2 = explicit, 1 = broader, 0 = peripheral).

The sum of these five criteria yielded a relevance score ranging from 0 (no relevance) to 10 (highly relevant). This rubric enabled a structured and reproducible comparison of the scientific relevance of results obtained through semantic search (POE) versus keyword-based search (GEO).

Manual relevance scoring was independently performed by two evaluators with expertise in bioinformatics, who rated each retrieved dataset on a 1–10 relevance scale according to its biological connection to the query.

Inter-rater agreement was assessed using Cohen’s κ (κ = 0.79), indicating substantial consistency between evaluators.

A two-tailed Wilcoxon signed-rank test comparing the total scores of the two experts confirmed that their ratings did not differ significantly (p = 0.097 > 0.05), demonstrating strong agreement and reliability in the manual assessments.

Consequently, the final relevance score for each dataset was calculated as the mean of both expert ratings.

To evaluate the significance of the observed differences between POE and GEO relevance scores across the benchmark queries, a two-tailed Wilcoxon signed-rank test was performed.

This non-parametric test was chosen for its robustness to non-normal distributions of ordinal data and revealed that POE achieved statistically higher relevance than GEO (p < 0.05).

These analyses enhance the methodological transparency and statistical rigor of the evaluation framework.

Embedding generation and vector representation

To support semantic search across the literature, POE incorporates a transformer-based embedding model implemented via the SentenceTransformers library [31] . Specifically, POE employs SBioBERT, a biomedical-specialized model pretrained on large-scale corpora including PubMed abstracts and PMC full-text articles, with additional sentence-level fine-tuning to capture domain-specific terminology and context more effectively. SBioBERT generates 768-dimensional dense vector embeddings optimized for biomedical semantic similarity ( Table 1 ). The model is available via Hugging Face: https://huggingface.co/dmis-lab/sentence-biobert .

SBioBERT was selected as the backbone embedding model of POE due to its strong performance in representing biomedical concepts, enabling accurate retrieval of semantically related literature and associated datasets. Through the web interface or RESTful API, users can seamlessly query the system using natural language, benefiting from SBioBERT’s ability to capture fine-grained biomedical context.

Before encoding, each PubMed title and abstract is concatenated, lowercased, stripped of non-informative characters, and truncated or padded to a maximum of 256 tokens. The resulting embeddings are stored and indexed using the FAISS library, enabling fast and meaningful retrieval of semantically similar content tailored to biomedical abstracts.

Semantic search and indexing

The 768-dimensional SBioBERT embeddings form the basis of POE’s semantic index, which is implemented using FAISS to enable fast and efficient retrieval. The index currently uses the IndexFlatL2 configuration, which performs exact nearest-neighbor search with perfect recall and precision, serving as the ground truth benchmark. While IndexFlatL2 provides the highest retrieval accuracy, it requires relatively higher memory usage and query time.

To evaluate trade-offs in performance, we benchmarked alternative FAISS index types, including IndexIVFFlat, IndexPQ, and IndexHNSW. IndexIVFFlat leverages an inverted file structure to reduce the search space, offering faster query times and lower memory requirements at the cost of approximate results. IndexPQ applies product quantization to compress vector representations, providing substantial memory and time efficiency but with a noticeable drop in retrieval quality. IndexHNSW employs a hierarchical navigable small-world graph that balances speed and accuracy, enabling efficient approximate search with moderate resources.

Although SBioBERT was originally optimized for sentence-level semantic understanding, in POE it is applied to variable-length biomedical text segments, including both PubMed abstracts and user-submitted queries.

To handle multi-sentence queries, all sentences are concatenated and processed as a single input sequence with a maximum token length of 256, allowing cross-sentence dependencies to be captured within a unified contextual window.

For multiword biomedical entities (e.g., Parkinson’s disease, oxidative stress response, transcription factor binding sites), SBioBERT’s WordPiece tokenizer decomposes complex terms into subword units while preserving their semantic integrity.

Embeddings corresponding to tokens belonging to the same entity or query are aggregated using mean pooling, generating a dense, context-aware representation that retains information about compositional biomedical terms.

This implementation enables POE to adapt SBioBERT beyond its original sentence-level scope, providing robust semantic encoding for realistic, multi-sentence biomedical queries encountered in literature-informed dataset retrieval.

The following table summarizes performance across these index types:

We systematically assessed these configurations by executing 100 representative semantic search queries derived from biomedical abstracts. For each query, the top 10 nearest neighbors were retrieved, and Recall@10 and Precision@10 were computed by comparing approximate methods against the IndexFlatL2 baseline. Additionally, average query time and memory usage were measured to reflect realistic usage scenarios.

As summarized in Table 2 , approximate methods offered substantial gains in speed and memory efficiency, but IndexFlatL2 consistently achieved the highest accuracy and most reliable semantic retrieval. Given the importance of precise dataset matching in biomedical discovery, IndexFlatL2 was selected as the default index type in POE.

All benchmarking was conducted on a machine with CPU Intel Core i3–10100 CPU @ 3.60 GHz, 8 cores, 64 GB RAM, using a dataset with more than 250,000 GEO-linked articles.

Backend infrastructure

POE's backend is implemented in Python using the FastAPI web framework. Metadata, records, and vector embeddings are stored in a PostgreSQL database with the pgvector extension enabled. All sensitive information, including user credentials and API tokens, is securely managed using hashed passwords.

The semantic search engine is integrated with the database and embedding pipeline, such that a user’s query is embedded into the same 768-dimensional space and used to identify semantically similar PubMed abstracts through FAISS. These results are then used to identify and rank associated GEO datasets. The entire backend service is deployed with unicorn, providing asynchronous I/O and fast response times for search queries and API endpoints.

Front-end and user interface

The POE user interface is implemented as a single-page web application using React ( https://react.dev/reference/react ), TypeScript [32] , and Tailwind CSS ( https://tailwindcss.com/docs/ ). It provides a responsive and accessible experience optimized for both desktop and mobile environments. The frontend communicates with the backend exclusively through RESTful API endpoints.

Users can submit natural language queries, browse semantically matched PubMed articles, and explore linked datasets. Structured filtering is available across multiple metadata fields, including organism, library strategy, experiment type, and publication year. Each result item is enriched with cross-links to external resources, including GEO, PubMed, and ENA, and may be exported in tabular format for downstream analysis. In addition, ENA FASTQ file links retrieved during data acquisition are presented directly in the interface, enabling users to access raw sequencing data alongside processed dataset metadata.

The current version of POE provides interactive result filtering, and summary statistics summarizing dataset counts, organism distribution, and experiment types retrieved per query.

It also features an interactive histogram of similarity scores, displaying the geometric distance (L2) of each retrieved result from the query embedding, allowing users to visually assess the semantic proximity of datasets and publications.

These elements collectively support intuitive data exploration and rapid evaluation of search relevance.

Collection basket for dataset management

POE includes an interactive “collection basket” feature that enables users to curate and temporarily store selected datasets during a search session. Datasets can be added to the basket from any search results page or filtered view. This functionality allows researchers to compile relevant datasets from multiple queries before initiating downloads or exporting metadata. The basket supports exporting the list of selected GEO or ENA accessions in standard formats, facilitating integration with external analysis pipelines or download managers.

Programmatic access via RESTful API and downstream analysis

In addition to the web-based user interface, POE provides a secure, token-based RESTful API that enables programmatic access to its semantic search capabilities and linked dataset records. All requests require a valid bearer token in the Authorization header; tokens expire after 24 h, and the API is rate-limited to one request every four seconds. Available endpoints include:

Semantic PubMed search, which uses FAISS-based vector similarity to retrieve publications relevant to a free-text query with optional metadata filters (organism, library strategy, experiment type, molecule, publication year).

Direct dataset/article lookup, allowing retrieval of metadata for specific GEO Series (GSE) accessions, PubMed IDs, DOIs, or article titles; and

ENA metadata download, which returns a tab-delimited file containing metadata and raw data file links for a given ENA project accession.

Python code examples are provided in the online documentation, enabling integration of POE’s capabilities into external pipelines, reproducible workflows, and automated dataset publication matching.

API documentation is available at https://github.com/DimitrisGrig/POE/wiki/API .

Beyond programmatic querying, POE results can be exported in standardized TSV or JSON formats, which preserve key metadata fields such as PubMed identifiers, GEO accessions, and ENA project links.

These outputs can be readily parsed in analytical environments such as R or Python using libraries like pandas for downstream filtering, aggregation, and study selection.

The retrieved dataset identifiers can then be used to fetch associated expression matrices or sequencing files through GEOquery, GEOparse, or ENA utilities, enabling subsequent analysis steps such as differential expression, meta-analysis, or data reprocessing.

In this way, POE acts as a semantic discovery layer that guides researchers toward relevant datasets, which can then be programmatically integrated into established analytical pipelines.

Tool and library versioning

POE was developed using Python 3.12.9. The backend employs FastAPI v0.115.11 for API management and Uvicorn v0.34.0 as the asynchronous web server. Semantic embeddings are generated using sentence-transformers v3.4.1 with the SBioBert model [33] . FAISS-cpu v1.10.0 is used for similarity indexing, and PostgreSQL v17.4 serves as the metadata database. The frontend interface is built with React v19.0.0, TypeScript v5.7.2, and Tailwind CSS v4.0.9, ensuring a modern and responsive user experience.

System update schedule

POE is updated daily via automated schedulers that orchestrate the end-to-end data retrieval, embedding, indexing, and deployment processes.

Results

Platform overview and dataset coverage

Public Omics Explorer (POE) is a fully operational web application that enables semantic exploration of omics datasets through their linked biomedical literature.As of September 2025, POE indexes a total of 262,173 GEO datasets (GSE records) and 9682,571 sample-level entries (GSM), integrating them with a collection of 130,548 PubMed publications. All GEO entries with valid PubMed identifiers are indexed with semantic embeddings, enabling vector-based similarity search from text queries to experimental datasets.

Users interact with the platform through a user-friendly interface (available at https://nplab.gr/poe ) that requires no installation or authentication ( Fig. 3 ). The POE web interface was iteratively refined through informal testing with graduate students and researchers in molecular biology and bioinformatics. User feedback emphasized clarity of search results, intuitive filtering controls, and smooth navigation between literature and dataset views.

These sessions confirmed that POE can be effectively used by non-computational scientists to explore datasets starting from natural-language queries.

Queries can be initiated using free-text natural language input, which is embedded into a shared semantic space alongside PubMed titles and abstracts. POE retrieves the most semantically similar articles, each linked to associated GEO datasets, enriched metadata, and direct links to GEO, PubMed, and ENA records.

At any point, users can refine their results through structured filters that include organism, sample type, library strategy, experiment type, publication year, and molecule extracted. Each result is displayed with summarized metadata and export options, allowing easy transition from conceptual queries to actionable datasets. Additionally, results can be further filtered based on a visually adjustable L2 similarity threshold.

Literature-informed dataset retrieval

Unlike conventional keyword-based search engines, POE supports concept-level discovery by leveraging the scientific context of publications. For example, a user query such as "inflammation in Parkinson’s disease" (see Supplementary Table 4 for POE result set) retrieves PubMed abstracts discussing neuroinflammatory pathways and neurodegeneration. From these, POE surfaces semantically associated GEO datasets, even when the metadata of those datasets do not explicitly mention either “Inflammation” or “Parkinson’s disease.”

This capability significantly enhances discovery by bridging the gap between implicit biological context in publications and structured experimental datasets. It enables researchers to identify datasets based on the relevance of the publication content, rather than reliance on exact metadata matches.

Comparative performance: keyword vs. semantic search

The comparative evaluation showed that POE and GEO returned overlapping but distinct sets of datasets. Notably, POE consistently retrieved substantially more results than GEO, reflecting the broader coverage enabled by its semantic indexing of PubMed-linked datasets. For instance, for “chromatin accessibility in stem cell fate decisions,” POE retrieved 506 datasets compared to just 6 PubMed-linked (7 total) datasets from GEO, while for “epigenetic control of lineage commitment,” POE returned 485 versus 10 from GEO. In several cases, GEO returned only a handful of results-for example, N = 2 in “miRNA promoter variants in cancer” and N = 1 for “pluripotency transcriptional networks” versus 784 and 116 from POE, respectively-underscoring the limitations of keyword-based matching for complex biological queries. This broader recall of POE is advantageous, as it expands the discovery space and reduces the risk of missing potentially relevant studies. In contrast, GEO’s keyword-based search is inherently constrained, since it requires all query terms to appear explicitly in the text of dataset annotations (unless more complex Boolean expressions with AND/OR are manually constructed, which substantially complicates the search process). As a result, many relevant datasets are not retrieved if even a single term is absent. A summary of the results for each query, including the number of POE results (total and with distance ≤1.0), GEO results with and without PubMed IDs, common GSE identifiers, and unique datasets per method, is provided in Supplementary Table 2 . The complete set of returned results for all queries across both systems is available in Supplementary_material_2 .

The qualitative relevance assessment was performed on the unique datasets (counts shown in the last two columns of Supplementary Table 2 ) returned by each system, focusing on queries for which both POE and GEO contributed at least four unique results. In total, 9 queries met this criterion, ensuring that the comparison was based on a sufficiently large and balanced set of records to yield statistically meaningful insights. The qualitative scoring highlighted systematic differences between POE-unique and GEO-unique results (see Supplementary Table 5 ). On average, POE-unique datasets achieved higher relevance scores than GEO-unique datasets, indicating that semantic search preferentially retrieved studies more closely aligned with the biological intent of the queries. For example, in the case study queries “chromatin accessibility in stem cell fate decisions” and “epigenetic control of lineage commitment,” POE uniquely identified datasets where the associated publications explicitly investigated the query concepts (mean score 7 out of 10), while GEO returned superficially matched records with only marginal or incidental relevance (mean score 2.3). Even in more challenging queries, such as “stress response pathways in cancer survival,” POE still maintained a higher mean score (6.1 vs 5.0). This trend was consistent across most queries, demonstrating that POE not only increases recall but also reduces noise by surfacing qualitatively more relevant datasets that are often missed by keyword-only retrieval.

This issue is exemplified by the query “inflammatory signaling in cardiovascular disease.” GEO returned a series of datasets unrelated to cardiovascular biology, such as “Molecular Clusters and Tumor-Immune Drivers of IgM Monoclonal Gammopathies” (focused on Waldenström macroglobulinemia), “Differential DNA Methylation Associated with Multiple Sclerosis in an Underrepresented Population” (neurological disease), and “The effect of rewarming ischemia on tissue transcriptome signatures: a clinical observational study in lung transplantation.” These records were captured primarily because of incidental overlaps in generic terms such as “inflammatory” or “signaling,” yet they lack any substantive connection to cardiovascular disease. Such mismatches illustrate the tendency of keyword-based retrieval in GEO to return biologically irrelevant datasets when queries involve complex, multi-word biological concepts.

In contrast, POE uniquely retrieved datasets directly aligned with the query’s intent. Representative examples include “Longitudinal profiling in patients undergoing cardiac surgery reveals postoperative changes in DNA methylation,” “Identification of a gene network driving the attenuated response to lipopolysaccharide of monocytes from hypertensive coronary artery disease patients,” and “Whole-Blood Transcriptome Unveils Altered Immune Response in Acute Myocardial Infarction Patients With Aortic Valve Sclerosis.” These studies explicitly investigate inflammatory processes in cardiovascular contexts-ranging from cardiac surgery and hypertension to acute myocardial infarction-thereby providing highly relevant biological insights that are missed by simple keyword filtering.

This trend was consistent across most queries, demonstrating that POE not only increases recall but also reduces noise by surfacing qualitatively more relevant datasets that are often missed by keyword-only retrieval.

To further illustrate this difference in practical terms, the following case example highlights how POE enables cross-context discovery and hypothesis generation beyond explicit keyword matches.

Case example: cross-context discovery of mechanistic links (miR-21 in breast cancer)

To illustrate this difference, the query “miR-21 breast cancer expression” demonstrates POE’s ability to uncover mechanistically related datasets beyond the explicitly annotated disease context. While the GEO keyword search retrieved directly annotated datasets such as GSE298584 ( “Fusobacterium nucleatum promotes metastasis of breast cancer via the miR-21–3p/FOXO3 axis” ), the POE semantic retrieval additionally surfaced functionally coherent studies including GSE109592 ( miR-21/DUSP8 axis in colorectal carcinoma ) and GSE117697 ( miR-21 in tumor-associated macrophages ).

Despite arising from distinct cancer models, these datasets converge on miR-21-mediated modulation of stromal and immune microenvironments, suggesting a shared pro-metastatic mechanism. This cross-context retrieval enabled the formulation of a new, data-driven hypothesis, that miR-21 may promote breast cancer metastasis through conserved pathways regulating immune and stromal cell signaling , similar to its role in other tumor types. Such examples highlight POE’s capability to facilitate hypothesis generation by integrating semantically linked datasets across biological systems that traditional keyword search fails to associate.

Comparative context with existing semantic systems

To further contextualize POE within existing semantic search systems, we compared it with representative biomedical literature-based tools, including PubTator3, LitSense, and BioASQ.

These systems enable entity recognition, sentence-level similarity search, or question answering within the biomedical literature but do not link retrieved articles to experimental datasets.

In contrast, POE uniquely supports dataset-level semantic retrieval by connecting PubMed publications to GEO and ENA records through SBioBERT embeddings and FAISS-based similarity indexing ( Table 3 ).

This functional distinction positions POE as a complementary resource to existing literature-centric systems, extending semantic discovery beyond publication text to the underlying omics data.

Additional access modes

Dataset search and article search

In addition to semantic queries, POE provides two direct access modes through the Dataset Search and Article Search menu options ( Fig. 4 ). The Dataset Search allows retrieval of datasets using GEO Series accessions (GSE IDs) or GEO titles, while the Article Search enables retrieval based on PubMed titles or DOIs. These modes provide rapid access to known studies without requiring exploratory queries. All retrieved records include direct cross-links to GEO for processed data, ENA/SRA for raw sequencing data, and PubMed for the corresponding publication.

Programmatic access

In addition to the web interface, POE offers a RESTful API that mirrors its core search and retrieval capabilities. This programmatic interface allows external tools, pipelines, and institutional platforms to query POE’s semantic index, retrieve matched datasets, and access linked records automatically. The API supports semantic PubMed search with filters, direct retrieval by identifiers (GSE ID, PubMed ID, DOI, title), and ENA metadata download for given project accessions. Token-based authentication ensures secure access, and Python examples are provided for integration into reproducible bioinformatics workflows.

Limitations and future work

While POE demonstrates strong performance in linking biomedical literature with omics datasets, several limitations highlight areas for future improvement.

First, the system currently focuses on GEO datasets that include PubMed-linked publications, which may introduce a coverage bias toward well-annotated or frequently cited studies.

To mitigate this limitation, future updates will explore automated strategies to infer probable publication links for additional GEO records lacking explicit PubMed identifiers.

These strategies will combine metadata alignment (e.g., title, author, organism, and submission date matching) with semantic similarity analysis between GEO study descriptions and PubMed abstracts using SBioBERT embeddings.

Second, POE currently embeds titles and abstracts from PubMed rather than full-text articles. This design reflects both licensing constraints and computational efficiency requirements. Nonetheless, the integration of full-text semantic embeddings, where legally and technically feasible, represents a valuable direction for enhancing contextual understanding and retrieval depth.

Third, the platform currently integrates transcriptomic datasets from GEO, while data from other omics domains such as proteomics (PRIDE) [34] and metabolomics (MetaboLights) [35] , are planned for inclusion. Extending POE to support cross-omics semantic search will further strengthen its role as a unifying discovery resource across diverse experimental modalities.

Finally, although the use of the IndexFlatL2 configuration in FAISS ensures exact nearest-neighbor retrieval and maximal accuracy, it is associated with higher memory and computational requirements compared to approximate methods. Future work will explore hybrid or hierarchical indexing strategies (e.g., IVFFlat or HNSW) to balance scalability with precision as the corpus continues to grow.

Collectively, these planned enhancements will expand POE’s coverage, improve retrieval efficiency, and extend its applicability across multiple omics domains, while maintaining transparency and reproducibility through its public API and open configuration pipeline.

Although POE was not specifically designed as a FAIR-compliance framework, its architecture inherently aligns with the FAIR principles of data management.

All indexed entities include persistent identifiers (GEO accessions, PubMed IDs, and ENA project IDs), ensuring that datasets remain findable and uniquely referenced.

The platform provides accessibility through both an intuitive graphical interface and a public REST API, enabling programmatic integration into external analytical workflows.

By maintaining standardized metadata structures compatible with major repositories, POE promotes interoperability, while direct linking to original data providers ensures reusability of omics datasets.

Collectively, these features position POE as a FAIR-aligned semantic search system that fosters transparent, accessible, and reusable biomedical data discovery.

In summary, POE embodies the principles of Findable, Accessible, Interoperable, and Reusable (FAIR) data management, ensuring long-term transparency and sustainability in biomedical data access.

While POE currently focuses on GEO-hosted datasets such as RNA-seq, ChIP-seq, and ATAC-seq, its semantic search architecture can be extended to other repositories hosting complementary data types, including those not yet fully integrated with GEO.

For instance, CAGE-seq datasets central to transcription start site analysis, such as those used in DeepTSS [ref] and AdaptCAGE [ref], are primarily hosted by the FANTOM consortium or in ArrayExpress.

In future iterations, POE’s framework could be adapted to include such repositories, enhancing its utility across a broader spectrum of regulatory genomics tools.

Similarly, literature-linked discovery of datasets relevant to microRNA regulation and promoter activity could facilitate deeper exploration using platforms such as miRGen v4 [ref]

Conclusions

POE introduces a transformative approach for discovering and reusing public omics data by semantically linking GEO datasets to PubMed literature. Unlike traditional keyword-based searches that often miss relevant studies due to inconsistent metadata or varying terminology, POE performs a semantic search across a subset of PubMed articles explicitly linked to GEO records. By embedding article titles and abstracts into a high-dimensional vector space, the platform enables the retrieval of conceptually relevant publications and their associated datasets using natural language queries.

This paradigm shift significantly lowers the barrier to data-driven discovery. Instead of requiring expertise in GEO metadata structure or precise keyword formulation, researchers can begin with a biological question-such as “inflammation in Parkinson’s disease”- and obtain relevant literature and datasets, even when dataset annotations do not explicitly contain those terms. This is particularly valuable in interdisciplinary contexts, where terminology often diverges across subfields. By leveraging the semantic content of publications as a proxy for dataset context, POE facilitates hypothesis generation, accelerates exploratory research, and supports meta-analyses, cross-condition comparisons, and validation studies. Furthermore, by employing the biomedical-optimized SBioBERT model - selected for its demonstrated superior performance in representing biomedical terminology and contextual relationships - POE adapts to the linguistic characteristics of biomedical literature, enabling accurate and context-aware retrieval for domain-specific searches.

A notable advantage of semantic search is its ability to accommodate queries expressed in natural language, closely resembling the way researchers formulate scientific questions. This makes it possible to retrieve datasets that align more precisely with the biological intent of the query. In contrast, keyword-based approaches typically require the construction of complex Boolean expressions with multiple AND and OR operators. Such queries can easily become overly restrictive-excluding relevant datasets overly permissive, yielding large sets of superficially matched results. This structural limitation of lexical search contributes to the lower qualitative relevance observed in GEO-unique results, whereas semantic search can capture the conceptual meaning of a query without relying on exhaustive keyword enumeration.

Beyond search functionality, POE promotes data reuse and reproducibility by facilitating the discovery of datasets aligned by conceptual relevance rather than rigid annotation. Its intuitive web interface and natural language support ensure accessibility across research backgrounds, enabling experimental biologists, clinicians, students, and educators to explore public omics data without requiring advanced computational skills. By lowering both technical and conceptual barriers, POE democratizes access to high-value omics resources and fosters broader participation in data-driven science and systems biology.

The strength of this approach was further substantiated by comparative evaluation. Across 30 benchmark queries, POE consistently retrieved substantially more datasets than GEO, in some cases exceeding GEO by two orders of magnitude. Crucially, this broader recall was not achieved at the expense of specificity: manual scoring of unique results revealed that POE-unique datasets were systematically more relevant to the biological intent of the queries compared to GEO-unique datasets. The case of “inflammatory signaling in cardiovascular disease” exemplifies this contrast, where GEO returned records unrelated to cardiovascular biology due to incidental keyword matches, while POE uniquely identified studies directly investigating inflammatory mechanisms in cardiac surgery, coronary artery disease, and myocardial infarction. These findings demonstrate that semantic indexing not only expands the discovery space but also improves the signal-to-noise ratio of retrieved datasets. By reducing irrelevant hits and surfacing contextually aligned results, POE provides a clear functional advantage over keyword-based search, supporting its role as a transformative tool for hypothesis-driven dataset discovery in biomedicine.

Funding

This work was developed and deployed using the computational infrastructure and servers of the Department of Informatics and Telecommunications, National and Kapodistrian University of Athens, where Nikos Perdikopanis is based. No dedicated research funding was received for the development of the Public Omics Explorer (POE).

Publication fees were covered by Georgios A. Pavlopoulos through the Hellenic Foundation for Research and Innovation (H.F.R.I.), under the call “Greece 2.0 - Basic Research Financing Action (Horizontal support of all Sciences), Sub-action II”, Grant ID: 16718-PRPFOR, and the “Greece 2.0 - National Recovery and Resilience Plan”, Grant ID: TAEDR-0539180.

Author Statement

The authors confirm that the following statements accurately describe the individual contributions, funding information, and potential conflicts of interest related to this work.

CRediT authorship contribution statement

Georgios A. Pavlopoulos: Writing – review & editing, Validation, Project administration, Funding acquisition. Antonis Giannakakis: Writing – review & editing, Validation. Dimitris Grigoriadis: Writing – review & editing, Software, Project administration, Formal analysis, Data curation, Conceptualization. Margaritis Tsifintaris: Writing – review & editing. Nikos Perdikopanis: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Resources, Project administration, Investigation.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

The authors would like to thank all collaborators from partner universities and research institutes for their valuable contributions to the development and validation of the Public Omics Explorer (POE).

Dimitris Grigoriadis would like to personally thank Afroditi Orfanidou for her invaluable moral support during the development of this work. A special and heartfelt acknowledgment goes to Kleoniki, whose quiet presence was a constant source of inspiration throughout this journey.

Footnotes

Supplementary data associated with this article can be found in the online version at doi:10.1016/j.csbj.2025.11.004 .

Contributor Information

Dimitris Grigoriadis, Email: d.grigoriadis@genekor.com.

Georgios A. Pavlopoulos, Email: pavlopoulos@fleming.gr.

Nikos Perdikopanis, Email: nikosp@uoa.gr.

Supplementary material

Supplementary material

Supplementary material

Supplementary material

Supplementary material

Supplementary material

Supplementary material

Supplementary material

Data availability

POE does not host or store raw omics data. Instead, it indexes structured metadata and provides direct external links to corresponding entries in GEO, PubMed, and ENA.

POE is a freely available web application, accessible at https://nplab.gr/poe .

References

1. Barrett T., Wilhite S.E., Ledoux P., Evangelista C., Kim I.F., Tomashevsky M., et al. NCBI GEO: archive for functional genomics data sets-update. Nucleic Acids Res. 2013 Jan 1;41(D1):D991–D995. doi: 10.1093/nar/gks1193.

2. Domrachev Edgar R., Lash M. AE. Gene Expression Omnibus: NCBI gene expression and hybridization array data repository. Nucleic Acids Res. 2002 Jan 1;30(1):207–210. doi: 10.1093/nar/30.1.207.

3. Wilkinson M.D., Dumontier M., IjJ Aalbersberg, Appleton G., Axton M., Baak A., et al. The FAIR Guiding Principles for scientific data management and stewardship. Sci Data. 2016 Mar 15;3(1) doi: 10.1038/sdata.2016.18.

4. Courtot M., Cherubin L., Faulconbridge A., Vaughan D., Green M., Richardson D., et al. BioSamples database: an updated sample metadata hub. Nucleic Acids Res. 2019 Jan;47(D1):D1172–D1178. doi: 10.1093/nar/gky1061.

5. Quiñones M., Liou D.T., Shyu C., Kim W., Vujkovic-Cvijin I., Belkaid Y., et al. METAGENOTE: a simplified web platform for metadata annotation of genomic samples and streamlined submission to NCBI’s sequence read archive. BMC Bioinforma. 2020 Sept;21(1):378. doi: 10.1186/s12859-020-03694-0.

6. Chua H.E., Tucker-Kellogg L., Bhowmick S.S. Proceedings of the 13th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics. ACM; New York, NY, USA: 2022. ArcheGEO.

7. Li Z., Li J., Yu P. GEOMetaCuration: a web-based application for accurate manual curation of Gene Expression Omnibus metadata. Database (Oxf) 2018 Jan;2018 doi: 10.1093/database/bay019.

8. Gonçalves R.S., O’Connor M.J., Martínez-Romero M., Egyedi A.L., Willrett D., Graybeal J., et al. CEDAR Work OntolAssist Environ Author metadata that Descr Sci Exp arXiv [csDB. 2019 May

9. Martínez Romero M. Proceedings of MOL2NET 2016, International Conference on Multidisciplinary Sciences, 2nd edition. MDPI; Basel, Switzerland: 2016. Using biomedical ontologies to improve metadata management in CEDAR project.

10. Bernstein M.N., Doan A., Dewey C.N. MetaSRA: normalized human sample-specific metadata for the Sequence Read Archive. Bioinformatics. 2017 Sept;33(18):2914–2923. doi: 10.1093/bioinformatics/btx334.

11. Giles C.B., Brown C.A., Ripperger M., Dennis Z., Roopnarinesingh X., Porter H., et al. ALE: automated label extraction from GEO metadata. BMC Bioinforma. 2017 Dec;18(S14) doi: 10.1186/s12859-017-1888-1.

12. Chen G., Ramírez J.C., Deng N., Qiu X., Wu C., Zheng W.J., et al. Restructured GEO: restructuring Gene Expression Omnibus metadata for genome dynamics analysis. Database (Oxf) 2019 Jan;2019 doi: 10.1093/database/bay145.

13. Mecham A., Stephenson A., Quinteros B.I., Brown G.S., Piccolo S.R. TidyGEO: preparing analysis-ready datasets from Gene Expression Omnibus. J Integr Bioinform. 2024 Mar;21(1) doi: 10.1515/jib-2023-0021.

14. Davis S., Meltzer P.S. GEOquery: a bridge between the Gene Expression Omnibus (GEO) and BioConductor. Bioinformatics. 2007 July;23(14):1846–1847. doi: 10.1093/bioinformatics/btm254.

15. Zhu Y., Davis S., Stephens R., Meltzer P.S., Chen Y. GEOmetadb: powerful alternative search engine for the gene expression omnibus. Bioinformatics. 2008 Dec;24(23):2798–2800. doi: 10.1093/bioinformatics/btn520.

16. Sozanska A.M., Fletcher C., Bihary D., Samarajiwa S.A. SpiderSeqR: an R package for crawling the web of high-throughput multi-omic data repositories for data-sets and annotation. bioRxiv. 2020 Apr

17. Cai L., Lin S., Zhou Y., Yang L., Ci B., Zhou Q., et al. Lung Cancer Explorer (LCE): an open web portal to explore gene expression and clinical associations in lung cancer. bioRxiv. 2018 Feb doi: 10.1038/s41388-018-0588-2.

18. Papatheodorou I., Fonseca N.A., Keays M., Tang Y.A., Barrera E., Bazant W., et al. Expression Atlas: gene and protein expression across multiple studies and organisms. Nucleic Acids Res. 2018 Jan;46(D1):D246–D251. doi: 10.1093/nar/gkx1158.

19. Stathias V., Turner J., Koleti A., Vidovic D., Cooper D., Fazel-Najafabadi M., et al. LINCS Data Portal 2.0: next generation access point for perturbation-response signatures. Nucleic Acids Res. 2020 Jan;48(D1):D431–D439. doi: 10.1093/nar/gkz1023.

20. Lachmann A., Torre D., Keenan A.B., Jagodnik K.M., Lee H.J., Wang L., et al. Massive mining of publicly available RNA-seq data from human and mouse. Nat Commun. 2018 Dec;9(1) doi: 10.1038/s41467-018-03751-6.

21. Perez-Riverol Y., Bai M., da Veiga Leprevost F., Squizzato S., Park Y.M., Haug K., et al. Discovering and linking public omics data sets using the Omics Discovery Index. Nat Biotechnol. 2017 May;35(5):406–409. doi: 10.1038/nbt.3790.

22. Piñero J., Ramírez-Anguita J.M., Saüch-Pitarch J., Ronzano F., Centeno E., Sanz F., et al. The DisGeNET knowledge platform for disease genomics: 2019 update. Nucleic Acids Res. 2020 Jan;48(D1):D845–D855. doi: 10.1093/nar/gkz1021.

23. Wei C.H., Allot A., Lai P.T., Leaman R., Tian S., Luo L., et al. PubTator 3.0: an AI-powered literature resource for unlocking biomedical knowledge. Nucleic Acids Res. 2024 July;52(W1):W540–W546. doi: 10.1093/nar/gkae235.

24. Yeganova L., Kim W., Tian S., Comeau D.C., Wilbur W.J., Lu Z. LitSense 2.0: AI-powered biomedical information retrieval with sentence and passage level knowledge discovery. Nucleic Acids Res. 2025 July;53(W1):W361–W368. doi: 10.1093/nar/gkaf417.

25. Tsatsaronis G., Balikas G., Malakasiotis P., Androutsopoulos I., Paliouras G., et al. An overview of the BIOASQ large-scale biomedical semantic indexing and question answering competition. BMC Bioinforma. 2015 Apr 22;16:138. doi: 10.1186/s12859-015-0564-6.

26. Sarkans U., Gostev M., Athar A., Behrangi E., Melnichuk O., Ali A., et al. The BioStudies database-one stop shop for all data supporting a life sciences study. Nucleic Acids Res. 2018 Jan;46(D1):D1266–D1270. doi: 10.1093/nar/gkx965.

27. Katayama T., Kawashima S., Okamoto S., Moriya Y., Chiba H., Naito Y., et al. TogoGenome/TogoStanza: modularized Semantic Web genome database. Database (Oxf) 2019 Jan;2019 doi: 10.1093/database/bay132.

28. Khomtchouk B.B., Dyomkin V., Vand K.A., Assimes T., Gozani O. Biochat: a database for natural language processing of Gene Expression Omnibus data. bioRxiv. 2018 Nov

29. Leinonen R., Akhtar R., Birney E., Bower L., Cerdeno-Tárraga A., Cheng Y., et al. The European Nucleotide Archive. Nucleic Acids Res. 2011 Jan;39:D28–D31. doi: 10.1093/nar/gkq967. (Database issue)

30. Beautiful Soup Documentation - Beautiful Soup 4.13.0 documentation [Internet]. [cited 2025 July 18]. Available from: 〈https://www.crummy.com/software/BeautifulSoup/bs4/doc/〉 .

31. Reimers N., Gurevych I. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks [Internet]. arXiv; 2019 [cited 2025 Aug 1]. Available from: 〈http://arxiv.org/abs/1908.10084〉 .

32. (a) Bierman G., Abadi M., Torgersen M. In: ECOOP 2014 - Object-Oriented Programming [Internet] Jones R., editor. Springer Berlin Heidelberg; Berlin, Heidelberg: 2014. Understanding TypeScript; pp. 257–281. [cited 2025 Sept 19] (b)(Hutchison D., Kanade T., Kittler J., Kleinberg J.M., Kobsa A., Mattern F., et al., editors. Lecture Notes in Computer Science; vol. 8586). Available from: http://link.springer.com/ 10.1007/978-3-662-44202-9_11 .

33. Thakur N., Reimers N., Daxenberger J., Gurevych I. Augmented SBERT: Data Augmentation Method for Improving Bi-Encoders for Pairwise Sentence Scoring Tasks [Internet]. arXiv; 2021 [cited 2025 Sept 19]. Available from: http://arxiv.org/abs/2010.08240.

34. Perez-Riverol Y., Bandla C., Kundu D.J., Kamatchinathan S., Bai J., Hewapathirana S., et al. The PRIDE database at 20 years: 2025 update. Nucleic Acids Res. 2025 Jan 6;53(D1):D543–D553. doi: 10.1093/nar/gkae1011.

35. Yurekten O., Payne T., Tejera N., Amaladoss F.X., Martin C., Williams M., et al. MetaboLights: open data repository for metabolomics. Nucleic Acids Res. 2024 Jan 5;52(D1):D640–D646. doi: 10.1093/nar/gkad1045.

Associated Data

Supplementary Materials

Supplementary material

Supplementary material

Supplementary material

Supplementary material

Supplementary material

Supplementary material

Supplementary material

Data Availability Statement

POE does not host or store raw omics data. Instead, it indexes structured metadata and provides direct external links to corresponding entries in GEO, PubMed, and ENA.

POE is a freely available web application, accessible at https://nplab.gr/poe .