Bridging the Omics Divide: A Modular Relational Approach to Multi-Layer Biological Data Management

1 Introduction

• Traditional file-based approaches struggle with large, evolving, multilayered datasets and can hinder integration, reuse, reproducibility, and collaborative analysis.Fragmented scripts and files can leave cross-layer insights and experimental results underused.Source: §1 · p. 2
• The central gap is a methodology combining competitive genomic querying with explicit modular multi-omics integration, rather than a lack of VCF storage or retrieval tools.The proposed framework links independently organized omics modalities through shared biological samples.Source: §1 · p. 3
• vcf2db models each omics modality as a distinct but linkable component centered on biological samples, with normalized genomic tables and an ingestion pipeline.The study evaluates 502 European 1000 Genomes samples and 25,057,597 annotated variants against Cutevariant, VCFdbR, and VCF-Miner across seven retrieval tasks.Source: §1 · p. 4
• The framework was extended with a transcriptomic layer without modifying genomic tables, linking modalities through a stable sample backbone.The study presents transcriptomic extension as a proof of concept for modular expansion.Source: §1 · pp. 4–5
• vcf2db performed strongly on selective coordinate and annotation queries, remained practically usable for genotype extraction, and exposed bottlenecks in aggregation-heavy workloads.The authors identify full-dataset aggregation and multi-level summarization as optimization targets.Source: §1 · p. 5

2 Methods

• The methods use VCF genomics as a proof of concept for a relational approach and describe its schema, transcriptomic extension, ingestion pipeline, and benchmark design.The design organizes biological observations around shared samples.Source: §2 · p. 5

2.1 The sample-centric design

• The sample-centric schema links sample metadata and genomic entities through keys, while normalization separates loci, alleles, annotations, and sample-specific genotypes.Chromosome partitioning and indexes on coordinates, annotation fields, and relational keys target common retrieval patterns.Source: §2.1 · pp. 6–8

2.2 Extensibility for multi-omics integration

• Additional modality-specific tables can reference the samples table, and the transcriptomic demonstration added gene-expression data without modifying existing genomic tables.This structure supports direct cross-layer retrieval while retaining independent organization for each modality.Source: §2.2 · pp. 8–9, Figure 2 · p. 8

2.3 Required Materials - Proof of concept

• The proof of concept used 25,057,597 VEP-annotated variants from 502 European 1000 Genomes samples and synthetic gene-expression data linked to those samples.The ingestion pipeline parses records iteratively and in batches; seven retrieval tasks were benchmarked against Cutevariant, VCFdbR, and VCF-Miner under controlled cache conditions.Source: §2.3 · pp. 9–11

3 Results

• The seven tasks are grouped into selective retrieval (Q1–Q3), genotype and variant-level queries (Q4–Q5), and aggregation-heavy workloads (Q6–Q7), with performance examined across cache states and dataset fractions.The analysis considers full-dataset timings, cache effects, and runtime scaling to reveal workload-specific trade-offs.Source: §3 · p. 11

3.1 Full-dataset benchmark performance across query classes

• Across query classes, vcf2db was strongest on lightweight retrieval but lost relative advantage on genotype and aggregation workloads.Source: §3.1 · pp. 11–12, Figure 3 · p. 12

3.1.1 Lightweight retrieval and annotation-driven filtering

• For Q1, vcf2db took 0.116 s cold and 0.050 s warm; for Q2, it led at 0.192 s cold and 0.097 s warm.Source: §3.1.1 · p. 12

3.1.2 Genotype retrieval and variant-level analytical queries

• For Q4, vcf2db took 8.502 s cold and 7.889 s warm; for Q5, it took 45.171 s cold and 43.828 s warm.Source: §3.1.2 · p. 13

3.1.3 Aggregation-heavy workloads

• Q6 took vcf2db 291.337 s cold versus 66.226 s for VCFdbR-file; Q7 took 124.751 s cold versus cutevariant’s 4.990 s warm.Source: §3.1.3 · p. 13

• Figure 4 compares warm-to-cold runtime ratios across queries and systems; ratios near 1 indicate computation-dominated, cache-insensitive performance.Source: §3.2 · pp. 13–14, Figure 4 · p. 14
• For vcf2db, Q1–Q3 ratios were 0.43, 0.51, and 0.44, while Q4, Q5, and Q7 were 0.93, 0.97, and 0.97.Source: §3.2 · p. 14
• Cutevariant’s ratios reached 0.02–0.09 on Q2, Q3, Q5, Q6, and Q7, while Q4 was 0.94; other systems showed query-dependent profiles.Source: §3.2 · pp. 14–15

Lightweight retrieval and annotation-driven filtering (Figure 5).

• For Q1–Q3, vcf2db scaled nearly flat; Q1 stayed below 0.2 s cold, while Q3 remained around 20.2–21.8 s cold across fractions.Source: Lightweight retrieval and… · p. 15, Figure 5 · p. 16

Genotype retrieval and variant-level analytical queries (Figure 6).

• Q4 grew from under 0.3 s to about 8.5 s cold as sample fraction rose, whereas vcf2db’s Q5 cold runtimes stayed near 45 s.Source: Genotype retrieval and variant-level… · pp. 16–17, Figure 6 · p. 17

Aggregation-heavy workloads (Figure 7).

• For Q6, vcf2db cold runtimes stayed at 310–316 s across fractions; Q7 remained at 122.7–138.0 s cold, keeping aggregation costly even at small cohorts.Source: Aggregation-heavy workloads (Figure 7) · pp. 17–18, Figure 7 · p. 18

3.3 Runtime scaling across dataset fractions

• Across workloads, variant-centric retrieval scaled negligibly with cohort size, sample-driven genotype extraction scaled near-linearly, and aggregation incurred high, largely sample-independent costs.Source: Aggregation-heavy workloads (Figure 7) · pp. 18–19

4.1 RQ1: Can annotated VCF data be represented relationally while preserving meaningful and efficient genomic retrieval?

• Across Q1–Q3, vcf2db achieved the lowest or near-lowest execution times, showing normalized entities can support targeted biological retrieval efficiently.The workloads covered genomic intervals, gene-consequence combinations, and rare high-impact variants; chromosome partitioning and targeted indexing supported retrieval.Source: §4.1 · p. 19
• Aggregation-heavy queries required broad scans, large joins, and summarization; their costs were largely independent of sample count and need further optimization.Proposed options include materialized views, precomputed summaries, denormalized analytical tables, and workload-specific indexes.Source: §4.1 · p. 20

4.2 RQ2: Can the schema be extended modularly beyond genomics?

• Adding gene-expression tables linked by shared sample identifiers required no changes to the existing genomic tables.The transcriptomic proof of concept used synthetic data, demonstrating incremental schema extension rather than biological discovery.Source: §4.2 · p. 20
• The schema keeps modality-specific organization while connecting molecular layers through shared samples, allowing cross-layer queries directly in the database.For example, samples carrying variants in a gene can be queried alongside expression values for the same or related genes.Source: §4.2 · p. 20

4.3 RQ3: Does the model remain competitive while providing a stronger basis for cross-layer querying?

• vcf2db was highly competitive on selective and annotation-driven queries, practical for genotype retrieval, but less efficient on aggregation-heavy workloads.The authors frame its contribution as combining competitive genomic retrieval with a schema designed for modular multi-omics integration, rather than replacing other systems in every scenario.Source: §4.3 · p. 21
• Evaluation used the European 1000 Genomes subset and synthetic transcriptomic data, and focused mainly on query runtime; larger real matched cohorts and operational metrics remain to be assessed.The authors also identify ingestion time, storage footprint, concurrent access, updates, interfaces, and downstream-pipeline integration as future evaluation areas.Source: §4.3 · pp. 21–22

5 Conclusion

• vcf2db decomposes annotated VCF data into normalized entities for variants, alternative alleles, annotations, genotypes, and samples, with explicit links to biological samples.Genomic variation served as the initial proof of concept because of its size, structure, annotation complexity, and sample-level dimensionality.Source: §5 · p. 23
• Selective coordinate- and annotation-driven queries performed strongly, genotype retrieval remained practical, and aggregation-heavy workloads remain the main optimization target.The conclusion reports these results across representative genomic retrieval workloads.Source: §5 · p. 23
• Transcriptomic data were added without redesigning the genomic component, with shared sample identifiers linking distinct omics modalities.The sample backbone allows each modality to retain its internal structure while supporting cross-layer database queries.Source: §5 · p. 23
• The proof of concept used the European 1000 Genomes subset, synthetic transcriptomic data, and primarily query-runtime evaluation.Further assessment should include larger and more diverse cohorts, real matched multi-omics data, and operational measures such as ingestion time and storage footprint.Source: §5 · p. 23
• The framework combines genomic retrieval with modular multi-omics organization, with future work targeting aggregation workloads, larger real matched cohorts, and additional molecular and clinical layers.The authors describe the sample-centered relational design as a way to query, connect, and reuse biological complexity.Source: §5 · pp. 23–24

Ethics approval and consent to participate

• The study involved no human participants or animals; its data were publicly available or synthetically generated.Source: Ethics approval and consent to… · p. 24

Availability of data and materials

• The study’s genomic data are available from the 1000 Genomes Project Phase 3 release, with results and ingestion code hosted on GitHub.Source: Availability of data and materials · p. 24

Competing interests

• The authors declared no competing interests.Source: Competing interests · p. 24

Funding

• The BioGeMT project at the University of Malta supported the work through a HORIZON-WIDERA-2022 grant.Source: Funding · p. 24

Authors’ contributions

• The authors’ roles covered study conception, methodology, software, validation, drafting, and manuscript finalization.Source: Authors’ contributions · p. 24