PhenotypeToGeneDownloaderR Pipeline
- PhenotypeToGeneDownloaderR is a pipeline for automated retrieval, harmonisation, and validation of phenotype-associated gene sets using diverse databases.
- It integrates 13 data sources—including curated, GWAS, and text-mined evidence—to capture complementary gene signals for robust analysis.
- Benchmark results demonstrate high recall (98.4%) and effective cross-source data merging, supporting applications like polygenic risk scoring and variant interpretation.
Searching arXiv for the specified paper and closely related work to ground the article. PhenotypeToGeneDownloaderR is a phenotype-guided R/Python pipeline for automated gene retrieval, harmonisation, symbol validation, and cross-source summary analysis. It is designed for workflows in which phenotype-associated gene sets serve as an upstream substrate for downstream prioritisation and interpretation, including polygenic risk score construction, enrichment testing, target prioritisation, and variant interpretation. Given a phenotype term, the pipeline queries integrated biological databases, standardises per-source outputs, combines gene lists, validates retrieved symbols against the NCBI human gene reference, and produces summary tables and visualisations. In the reported benchmark across 13 clinically relevant phenotypes and 13 databases, it generated 136,487 raw gene retrievals, retained 100,175 of 114,345 combined input symbols after validation, and recovered 1,039 of 1,056 genes in an HPO/ClinVar/OMIM-derived gold standard, corresponding to 98.4% recall (Muneeb et al., 2 May 2026).
1. Definition, scope, and workflow
PhenotypeToGeneDownloaderR addresses a recurrent problem in phenotype-first genomics: relevant evidence is distributed across heterogeneous databases with different interfaces, formats, scopes, and evidence models. The pipeline is therefore organised as an end-to-end retrieval and validation system rather than as a single predictive model. Its function is to assemble candidate gene sets from multiple evidence channels, retain source provenance, and deliver a validated and traceable output suitable for downstream ranking, enrichment, or clinical review (Muneeb et al., 2 May 2026).
The workflow begins from a plain-language phenotype term such as “Hypertension.” Source-specific modules query multiple integrated databases, and each module returns a source-native set of candidate genes for the phenotype. Duplicates are removed within each source and phenotype before further processing. Minimal harmonisation is then applied to preserve provenance and evidence type while aligning outputs to a common gene-symbol field. Per-source lists are unioned into a combined all-source set, which is then validated against the NCBI Homo_sapiens.gene_info reference. The validated output is accompanied by per-source and per-phenotype summaries, cross-source overlap statistics, and files segregating valid and invalid symbols (Muneeb et al., 2 May 2026).
This architecture places the tool upstream of prioritisation and interpretation rather than making it itself a causal inference engine. A plausible implication is that its principal methodological contribution lies in evidence aggregation and identifier normalisation, not in phenotype-to-gene representation learning.
2. Integrated databases and evidence model
The benchmark integrates 13 databases spanning curated clinical, GWAS, pathway, text-mined, protein-interaction, and expression evidence. This breadth is central to the pipeline’s design because the reported cross-source overlap is low, indicating that single-source retrieval omits substantial complementary signal (Muneeb et al., 2 May 2026).
| Database | Evidence type or role |
|---|---|
| Open Targets | Integrative disease–target platform; multi-evidence associations |
| GWAS Catalog | GWAS loci mapped to genes; association evidence |
| OMIM | Curated Mendelian disease–gene relationships |
| PubMed | Literature-derived gene mentions linked to the phenotype |
| GTEx | Gene–phenotype links via expression/tissue relevance modules |
| ClinVar | Clinical variant–gene evidence curated from submissions |
| Human Phenotype Ontology (HPO) | Phenotype–gene annotations curated from clinical/phenotypic literature |
| UniProt | Protein knowledgebase; phenotype/disease annotations and cross-references |
| Reactome | Pathway memberships enriched for phenotype terms |
| STRING-DB | Protein–protein interactions; indirect disease relevance via network context |
| Gene Ontology (GO) | Functional annotations matching phenotype terms |
| KEGG | Pathway memberships |
| DisGeNET | Disease–gene associations |
The source modules encapsulate source-specific access mechanisms, including APIs and parsers. GWAS Catalog retrieval is implemented via the gwasrapidd R package. DisGeNET access is API-dependent, and in the reported benchmark it yielded no output because access was unavailable. KEGG also produced zero validated genes in that benchmark, not because of an asserted absence of relevant evidence, but because compound gene-name strings in the analysed KEGG outputs were not split by the validation parser and were therefore marked invalid (Muneeb et al., 2 May 2026).
The evidence model is intentionally heterogeneous. Open Targets aggregates multiple evidence classes; HPO, OMIM, and ClinVar contribute curated phenotype and disease knowledge; PubMed adds text-mined associations; STRING-DB contributes network context; and GO, KEGG, and Reactome contribute functional or pathway structure. This suggests that the pipeline is designed to maximise retrieval breadth before any downstream evidence weighting is applied.
3. Harmonisation schema and symbol validation
A central component of PhenotypeToGeneDownloaderR is its harmonisation and validation layer. The per-source harmonised record includes the phenotype term, source, the raw gene symbol returned by the source, a high-level evidence type, and optional source-specific fields retained for traceability. During validation, additional fields are appended: symbol_validated, gene_id, validation_status, and synonym_used (Muneeb et al., 2 May 2026).
Validation is performed against the NCBI Homo_sapiens.gene_info reference, downloaded in April 2026, including official symbols, Entrez GeneIDs, HGNC-approved names, and the NCBI synonym field. The procedure has three stages: direct match of symbol_raw to the official NCBI human symbol list; synonym-based mapping via the NCBI synonym field if direct matching fails; and invalidation if neither route succeeds. The benchmark explicitly quantifies the contribution of synonym handling: 4,912 of 100,175 validated symbols, or 4.9%, were retained through synonym rescue rather than direct matching (Muneeb et al., 2 May 2026).
The reported formulas are:
and
The output schema supports provenance-preserving downstream analysis. Typical columns in the combined validated outputs are phenotype, source, symbol_raw, symbol_validated, gene_id, validation_status, synonym_used, evidence_type, and source-specific fields retained for traceability. Valid and invalid files are segregated by validation_status (Muneeb et al., 2 May 2026).
This validation framework addresses a common source of irreproducibility in gene retrieval pipelines: symbol drift, aliasing, and cross-database naming inconsistency. The reported synonym rescue rate indicates that direct symbol matching alone would have discarded a nontrivial fraction of otherwise recoverable genes.
4. Benchmark results and empirical performance
The benchmark covers 13 clinically relevant phenotypes and 13 databases. Summed across all phenotype–source cells, the pipeline produced 136,487 raw gene retrievals. After cross-source merging, 114,345 combined unique input symbols were subjected to validation, of which 100,175 were retained and 14,170 were marked invalid, for an overall validation rate of 87.6% (Muneeb et al., 2 May 2026).
Evaluation against a curated gold standard derived from HPO, ClinVar, and OMIM used a conservative rule: a gene entered the reference set for a phenotype if it appeared in at least two of those three curated databases. The resulting gold standard contained 1,056 genes across phenotypes, of which 1,039 were recovered somewhere in the validated output, yielding 98.4% recall. Full recall was achieved for 8 phenotypes, including asthma, BMI, GERD, allergic rhinitis, high cholesterol, hypertension, IBS, and osteoarthritis, while migraine recall was 83.0% (Muneeb et al., 2 May 2026).
Per-phenotype validation examples illustrate the range of performance. Hypertension yielded 15,320 combined raw symbols and 14,132 validated symbols, corresponding to a 92.2% validation rate. Body mass index yielded 23,361 raw symbols and 22,168 validated symbols, corresponding to 94.9%. Asthma yielded 14,524 raw and 12,919 validated symbols, corresponding to 88.9%. Cholesterol-lowering medication yielded 1,821 raw and 1,168 validated symbols, corresponding to 64.1% (Muneeb et al., 2 May 2026).
The hypertension case also makes the support distribution explicit. Among validated hypertension genes, 10,784 were supported by exactly 1 source, 2,040 by 2 sources, 409 by 3 sources, 117 by 4 sources, and 75 by 5 or more sources; 4.5% were supported by at least 3 sources. Hypertension achieved 100.0% gold-standard recall, recovering 258 of 258 curated genes, and had Precision@20 of 90.0% (Muneeb et al., 2 May 2026).
These figures indicate high recall with substantial evidence heterogeneity. A plausible implication is that the pipeline is tuned for comprehensive candidate-set construction rather than aggressive precision filtering at the retrieval stage.
5. Cross-source complementarity and comparative position
Cross-source overlap was reported to be low. Using the Jaccard index,
the mean non-zero pairwise Jaccard across phenotypes where both sources returned results was 0.026, and the maximum observed value was 0.106, highest between UniProt and GO. Most genes per phenotype were supported by a single source, while the fraction supported by 3 or more sources ranged from 0.2% for BMI to 4.5% for hypertension (Muneeb et al., 2 May 2026).
Unique-only contributions further quantify complementarity. Across all phenotypes, 35,174 of 41,017 GWAS Catalog genes were unique to GWAS, 29,504 of 40,916 Open Targets genes were unique to Open Targets, 7,493 of 11,706 OMIM genes were unique to OMIM, 4,124 of 5,209 GTEx genes were unique to GTEx, 1,817 of 2,776 ClinVar genes were unique to ClinVar, and 258 of 404 Reactome genes were unique to Reactome. These values support the paper’s conclusion that integrated evidence sources are complementary rather than redundant (Muneeb et al., 2 May 2026).
In module-level comparison, the full integrated pipeline recovered 1,039 of 1,056 gold-standard genes, corresponding to 98.4% recall. The GWAS Catalog module alone recovered 155 of 1,056, corresponding to 14.7% recall, and returned results for 8 of 13 phenotypes. The Open Targets module recovered 705 of 1,056, corresponding to 66.8% recall, and returned results for 11 of 13 phenotypes. The g:Profiler input consisting of the top 500 source-ranked validated genes per phenotype recovered 793 of 1,056, corresponding to 75.1% recall; this was explicitly noted as measuring retention of known genes in top-ranked sets rather than g:Profiler as a retrieval tool (Muneeb et al., 2 May 2026).
A related but methodologically distinct direction is KG-driven phenotype-to-gene ranking. PhenoKG predicts likely causative genes for rare monogenic diseases directly from patient phenotypes, with or without an available candidate list, using a rare disease knowledge graph together with GATv2 and transformer-based encoding. On MyGene2, it achieved 24.64% MRR and 33.64% nDCG@100, compared with 19.02% MRR and 30.54% nDCG@100 for SHEPHERD (Zaripova et al., 16 Jun 2025). The distinction is important: PhenotypeToGeneDownloaderR is an upstream retrieval, harmonisation, and validation framework, whereas PhenoKG is a phenotype-to-gene ranking model. Confusing the two would obscure their complementary roles in a diagnostic or prioritisation workflow.
6. Implementation, outputs, limitations, and future directions
The workflow is implemented primarily in R with source-specific modules, is released under the MIT licence, and is available at https://github.com/MuhammadMuneeb007/PhenotypeToGeneDownloaderR. It uses gwasrapidd for GWAS Catalog retrieval and g:Profiler for enrichment analysis of the top 500 validated, source-ranked genes per phenotype where available. Runtime was measured via /usr/bin/time, with MaxRSS captured; the total benchmark runtime was 94.11 minutes over 13 phenotypes, the mean runtime was 7.24 minutes per phenotype, and peak RAM was approximately 0.62–0.67 GB with mean 0.641 GB (Muneeb et al., 2 May 2026).
For each phenotype, the pipeline produces per-source gene counts and success status, per-database retrieval performance summaries, combined all-source gene lists before validation, validated and invalid files after validation, validated counts per phenotype and source, synonym-rescue summaries, pairwise Jaccard similarity matrices, gene-support frequency distributions, unique-only gene contribution summaries, gold-standard construction and evaluation tables, and enrichment outputs. The reported enrichment stage identified 41,787 significant terms across all phenotypes at FDR (Muneeb et al., 2 May 2026).
Several limitations are explicit. Phenotype term ambiguity can suppress retrieval when source modules rely on exact term matching, particularly in pathway and ontology resources. KEGG parsing failed in the benchmark because compound gene-name fields were not split by the validation parser. DisGeNET output was absent because API access was unavailable during the runs. STRING-DB is a protein-interaction resolver rather than a disease database, so its outputs may be sparse or indirect. The master coordination script also had a file-detection issue that caused some GWAS runs to be marked as failed despite successful retrieval, underscoring the importance of robust output detection (Muneeb et al., 2 May 2026).
The paper’s future directions are correspondingly infrastructural and methodological: hardening source connectors, improving KEGG parsing and DisGeNET integration, expanding identifier support beyond the current validation regime, improving disambiguation for gene families and homonyms, introducing evidence-weighted ranking, integrating variant-level data and network-based propagation, and adding ontology-aware retrieval that maps user phenotype terms automatically to HPO or OMIM. These proposals imply an evolution from retrieval-and-validation toward more explicit evidence weighting while preserving the tool’s current role as a lightweight, reproducible upstream framework (Muneeb et al., 2 May 2026).
