Sources of evidence

Genome-wide association studies (GWAS) ↑

gwas_credible_sets · Open Targets / Gentropy · weight ×1.0 · 946,743 links · 18,835 genes · 10,465 diseases · homepage

Genotype millions of common variants across tens or hundreds of thousands of people, then ask which variants occur more often in those with the disease. This is a hypothesis-free scan of the whole genome.

The complication is that the variant you detect is usually not the one doing the work. Nearby variants are inherited together in blocks, so a signal implicates a whole neighbourhood, and around 90% of these signals fall outside protein-coding sequence — in switches that regulate genes, sometimes genes a long way off. Two further steps are therefore applied: fine-mapping, which narrows the block to a "credible set" of variants statistically likely to contain the causal one, and locus-to-gene assignment, a machine-learning model that picks the gene most likely to be affected, using distance, whether the variant is known to alter that gene's expression, and 3-D contacts between the switch and the gene.

Limitations: the gene assignment is a prediction, not an observation. GWAS also finds common variants of small effect, so a strong association can mean a real but tiny contribution to risk.

• Mountjoy E, et al. An open approach to systematically prioritize causal variants and genes at all published human GWAS trait-associated loci. Nature Genetics 53:1527–1533 (2021). doi.org/10.1038/s41588-021-00945-5
• Ghoussaini M, et al. Open Targets Genetics: systematic identification of trait-associated genes using large-scale genetics and functional genomics. Nucleic Acids Research 49:D1311–D1320 (2021). doi.org/10.1093/nar/gkaa840

ClinVar — clinical variant interpretation ↑

eva · NIH/NCBI ClinVar via EMBL-EBI EVA · weight ×1.0 · 39,626 links · 7,315 genes · 9,428 diseases · homepage

When a diagnostic laboratory sequences a patient, it interprets each variant it finds and submits that interpretation to ClinVar: pathogenic, benign, or uncertain, with the condition it relates to. An association here means a variant in this gene has been judged to cause this disease in real patients.

Limitations: submissions vary in rigour, laboratories sometimes disagree about the same variant, and the archive reflects what gets tested — heavily weighted toward rare paediatric disease and hereditary cancer.

• Landrum MJ, et al. ClinVar: improvements to accessing data. Nucleic Acids Research 48:D835–D844 (2020). doi.org/10.1093/nar/gkz972
• Cezard T, et al. The European Variation Archive: a FAIR resource of genomic variation for all species. Nucleic Acids Research 50:D1216–D1220 (2022). doi.org/10.1093/nar/gkab960

Genomics England PanelApp ↑

genomics_england · Genomics England · weight ×1.0 · 13,190 links · 4,393 genes · 6,731 diseases · homepage

Expert panels decide which genes belong on a diagnostic test for a given rare disease, rating each gene green, amber or red by strength of evidence. Green means the panel would act on a variant in that gene clinically.

Limitations: coverage follows the diseases the UK 100,000 Genomes Project set out to diagnose. Silence about a gene is not a judgement against it.

• Martin AR, et al. PanelApp crowdsources expert knowledge to establish consensus diagnostic gene panels. Nature Genetics 51:1560–1565 (2019). doi.org/10.1038/s41588-019-0528-2

Rare-variant burden testing ↑

gene_burden · Multiple (Regeneron, AstraZeneca, Genebass, FinnGen, others) · weight ×1.0 · 8,454 links · 2,091 genes · 1,240 diseases · homepage

A single rare variant is seen too few times to reach statistical significance on its own. Burden testing instead collapses all the rare, likely-damaging variants in a gene into one count per person, and asks whether people with the disease carry more of them than people without. It trades the ability to name a variant for the power to name a gene.

Limitations: the answer depends on which variants you decided were "likely damaging" — a judgement made by a prediction algorithm before the test is run.

• Backman JD, et al. Exome sequencing and analysis of 454,787 UK Biobank participants. Nature 599:628–634 (2021). doi.org/10.1038/s41586-021-04103-z
• Karczewski KJ, et al. Systematic single-variant and gene-based association testing of thousands of phenotypes in 394,841 UK Biobank exomes (Genebass). Cell Genomics 2:100168 (2022). doi.org/10.1016/j.xgen.2022.100168
• Wang Q, et al. Rare variant contribution to human disease in 281,104 UK Biobank exomes. Nature 597:527–532 (2021). doi.org/10.1038/s41586-021-03855-y

Orphanet ↑

orphanet · Orphanet · weight ×1.0 · 7,412 links · 3,955 genes · 3,769 diseases · homepage

A curated reference for rare disease, in which experts record the genes known to cause each disorder and the mode of inheritance.

Limitations: rare disease only, by design.

• Orphanet: an online database of rare diseases and orphan drugs. INSERM, 1997– . Accessed via Open Targets 26.06. www.orpha.net/

UniProt — curated from literature ↑

uniprot_literature · UniProt Consortium · weight ×1.0 · 6,584 links · 4,200 genes · 5,112 diseases · homepage

A curator read the papers and recorded a gene–disease relationship. Note this is not text-mining despite the name: the claim is a human expert's, and the literature is the evidence they cite. That is why it carries full weight and sits under human genetics, while Europe PMC does not.

Limitations: curation lags publication, sometimes by years.

• The UniProt Consortium. UniProt: the Universal Protein Knowledgebase in 2025. Nucleic Acids Research 53:D609–D617 (2025). doi.org/10.1093/nar/gkae1010

UniProt — curated disease variants ↑

uniprot_variants · UniProt Consortium · weight ×1.0 · 5,127 links · 3,454 genes · 4,228 diseases · homepage

UniProt curators record specific amino-acid changes reported to cause disease, attached to the protein they alter.

Limitations: a variant reported once in one patient and a variant seen in hundreds both appear as entries.

• The UniProt Consortium. UniProt: the Universal Protein Knowledgebase in 2025. Nucleic Acids Research 53:D609–D617 (2025). doi.org/10.1093/nar/gkae1010

Gene2Phenotype ↑

gene2phenotype · EMBL-EBI · weight ×1.0 · 4,936 links · 2,853 genes · 3,250 diseases · homepage

Curated gene panels for developmental disorders, and for eye, skin and cardiac conditions, each entry rated by confidence and annotated with the mechanism — whether disease results from losing the protein, from gaining a new harmful activity, or from having only one working copy.

Limitations: restricted to the panels curated so far.

• Thormann A, et al. Flexible and scalable diagnostic filtering of genomic variants using G2P with Ensembl VEP. Nature Communications 10:2373 (2019). doi.org/10.1038/s41467-019-10016-3

ClinGen gene–disease validity ↑

clingen · Clinical Genome Resource · weight ×1.0 · 4,452 links · 2,887 genes · 2,219 diseases · homepage

Expert panels apply a published framework to score each gene–disease pair on the accumulated genetic and experimental evidence, arriving at a verdict from Definitive down through Limited to Disputed or Refuted. It is the closest thing the field has to a formal verdict on whether a gene really causes a disease.

Limitations: deliberately slow. Only a few thousand pairs have been adjudicated.

• Strande NT, et al. Evaluating the clinical validity of gene–disease associations: an evidence-based framework developed by the Clinical Genome Resource. American Journal of Human Genetics 100:895–906 (2017). doi.org/10.1016/j.ajhg.2017.04.015
• Rehm HL, et al. ClinGen — the Clinical Genome Resource. New England Journal of Medicine 372:2235–2242 (2015). doi.org/10.1056/NEJMsr1406261

COSMIC Cancer Gene Census ↑

cancer_gene_census · Wellcome Sanger Institute / COSMIC · weight ×1.0 · 91,572 links · 749 genes · 1,048 diseases · homepage

A manually curated catalogue of genes with mutations causally implicated in cancer, each classified by the mechanism — whether the gene normally restrains growth and is lost, or normally drives growth and is overactivated — and by the tumour types involved.

Limitations: a gene is listed for cancer broadly and for its specific tumour types, so common cancers inherit long gene lists.

• Sondka Z, et al. The COSMIC Cancer Gene Census: describing genetic dysfunction across all human cancers. Nature Reviews Cancer 18:696–705 (2018). doi.org/10.1038/s41568-018-0060-1
• Tate JG, et al. COSMIC: the Catalogue Of Somatic Mutations In Cancer. Nucleic Acids Research 47:D941–D947 (2019). doi.org/10.1093/nar/gky1015

IntOGen — computational driver detection ↑

intogen · Barcelona Biomedical Genomics Lab · weight ×1.0 · 2,595 links · 630 genes · 67 diseases · homepage

Tumours accumulate far more mutations than matter. IntOGen analyses thousands of sequenced tumours and looks for genes mutated more than the local background rate predicts, or whose mutations cluster in one part of the protein, or fall disproportionately on positions likely to change function. Those signatures distinguish drivers from passengers.

Limitations: it is a statistical inference over cohorts, so it finds recurrent drivers and misses ones specific to rare subtypes.

• Martínez-Jiménez F, et al. A compendium of mutational cancer driver genes. Nature Reviews Cancer 20:555–572 (2020). doi.org/10.1038/s41568-020-0290-x

ClinVar — somatic ↑

eva_somatic · NIH/NCBI ClinVar via EMBL-EBI EVA · weight ×1.0 · 2,246 links · 1,106 genes · 436 diseases · homepage

The same archive as ClinVar germline, restricted to variants acquired in tumour tissue rather than inherited.

Limitations: far smaller and less consistently submitted than the germline half.

• Landrum MJ, et al. ClinVar: improvements to accessing data. Nucleic Acids Research 48:D835–D844 (2020). doi.org/10.1093/nar/gkz972

ChEMBL clinical precedence ↑

clinical_precedence · Open Targets Clinical Mining · weight ×1.0 · 107,593 links · 1,516 genes · 3,233 diseases · homepage

Drug databases record what each drug binds to and what it is approved or trialled for. Joining the two gives: this protein is the target of a drug used against this disease. If blocking a protein treats a disease, the protein is presumably part of it.

Limitations — and this one matters more than any other caveat on this page: the reasoning is partly circular. A drug exists because someone already believed the target mattered, so this evidence partly measures the field's prior beliefs and its funding history rather than independent biology. It also attributes the whole effect to the known target, when many drugs act through additional or unknown mechanisms. Ask "what causes this disease?" and you should probably switch this off; ask "what is known about this disease?" and it belongs.

A concrete symptom: the strongest gene for many inflammatory diseases becomes NR3C1, the receptor that steroids act on. Steroids are prescribed across all of medicine — that is pharmacology leaking into a measure meant to describe disease biology.

• Zdrazil B, et al. The ChEMBL Database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods. Nucleic Acids Research 52:D1180–D1192 (2024). doi.org/10.1093/nar/gkad1004

CRISPR screens ↑

crispr_screen · CRISPRbrain · weight ×1.0 · 10,954 links · 7,286 genes · 7 diseases · homepage

CRISPR is a molecular tool that cuts DNA at a location specified by a short guide RNA, disabling the gene there. A screen industrialises this: build a library of guide RNAs — one or more for every gene in the genome — put it into a population of cells so that each cell is knocked out for a different single gene, then apply a challenge and see which cells thrive or die.

Because every guide is a DNA barcode, you do not need to track cells individually. Sequence the guides in the population before and after, and any guide that became rarer marks a gene the cells needed; any that became commoner marks a gene that was holding them back.

Turning that into disease evidence needs a disease-relevant readout. These screens are run in human cells differentiated into a relevant type — neurons, for instance — and the measured phenotype is chosen to reflect the disease process: survival under oxidative stress, accumulation of a protein that aggregates in neurodegeneration, or a specific stress response. A gene whose loss changes that phenotype is implicated in the process.

Limitations: a cell in a dish is not a person. The readout is a proxy chosen by the experimenter, so the evidence is only as good as the assumption that it stands for the disease. Note the coverage in the table below: thousands of genes but a handful of diseases, because each screen interrogates every gene against one carefully constructed phenotype.

• Tian R, et al. CRISPR interference-based platform for multimodal genetic screens in human iPSC-derived neurons. Neuron 104:239–255 (2019). doi.org/10.1016/j.neuron.2019.07.014
• CRISPRbrain: a data commons for functional genomics screens in differentiated human cell types. crisprbrain.org/

Reactome ↑

reactome · Reactome Collaboration · weight ×1.0 · 2,932 links · 1,652 genes · 218 diseases · homepage

An expert-curated map of biological reactions — which molecule acts on which, in what order. Some reactions are annotated as disrupted in a specific disease, which links every gene in that reaction to the disease.

Limitations: the link is to a pathway, not to a patient. It says the gene participates in a process known to go wrong, not that this gene's variation causes the disease.

• Milacic M, et al. The Reactome Pathway Knowledgebase 2024. Nucleic Acids Research 52:D672–D678 (2024). doi.org/10.1093/nar/gkad1025

Project Score — cancer dependencies ↑

crispr · Wellcome Sanger Institute Project Score · weight ×1.0 · 517 links · 302 genes · 26 diseases · homepage

The same CRISPR screening technique, aimed at a different question. Knock out every gene in turn across hundreds of different cancer cell lines and record which knockouts kill which lines. A gene that a particular cancer cannot survive without is a dependency — and a candidate drug target, since a drug blocking it would kill those cells and spare others.

Limitations: cell lines drift from the tumours they came from, and dependency in culture does not guarantee dependency in a patient.

• Behan FM, et al. Prioritization of cancer therapeutic targets using CRISPR–Cas9 screens. Nature 568:511–516 (2019). doi.org/10.1038/s41586-019-1103-9
• Dwane L, et al. Project Score database: a resource for investigating cancer cell dependencies and prioritizing therapeutic targets. Nucleic Acids Research 49:D1365–D1372 (2021). doi.org/10.1093/nar/gkaa882

Cancer Biomarkers ↑

cancer_biomarkers · Cancer Genome Interpreter · weight ×0.5 · 469 links · 184 genes · 77 diseases · homepage

Curated genomic markers that predict whether a tumour will respond to a particular drug — a mutation that makes a cancer treatable, or resistant.

Limitations: small, and about drug response rather than about what caused the cancer. Down-weighted to 0.5 by Open Targets.

• Tamborero D, et al. Cancer Genome Interpreter annotates the biological and clinical relevance of tumor alterations. Genome Medicine 10:25 (2018). doi.org/10.1186/s13073-018-0531-8

International Mouse Phenotyping Consortium ↑

impc · International Mouse Phenotyping Consortium · weight ×0.2 · 3,968,761 links · 10,856 genes · 9,406 diseases · homepage

A systematic effort to disable every protein-coding gene in the mouse, one at a time, and put the resulting animals through a standardised battery of measurements — body composition, blood chemistry, behaviour, hearing, vision, skeleton and more. The phenotypes observed are recorded in a controlled vocabulary and then mapped onto the human diseases they resemble.

This is genuine experimental causation: the gene was broken deliberately and something changed. It is also, by volume, the largest single source here.

Limitations: the mouse is not a small human. Many genes produce no phenotype in a clean laboratory background, some produce a different one, and the mapping from mouse phenotype to human disease is lossy — which is why Open Targets down-weights it to 0.2 despite its volume.

• Groza T, et al. The International Mouse Phenotyping Consortium: comprehensive knockout phenotyping underpinning the study of human disease. Nucleic Acids Research 51:D1038–D1045 (2023). doi.org/10.1093/nar/gkac972
• Dickinson ME, et al. High-throughput discovery of novel developmental phenotypes. Nature 537:508–514 (2016). doi.org/10.1038/nature19356

Expression Atlas ↑

expression_atlas · EMBL-EBI · weight ×0.2 · 167,346 links · 24,864 genes · 213 diseases · homepage

Measure how much RNA each gene produces in diseased tissue and in healthy tissue, and report the genes that differ.

Limitations: this is correlation, and the direction of causation is genuinely unknown. A gene may be up because it drives the disease, up because the body is responding to the disease, or up because the diseased sample contains more of a cell type that happens to express it. Note that it touches nearly every gene but very few diseases — a small number of comparisons applied genome-wide.

• George N, et al. Expression Atlas update: insights from sequencing data at both bulk and single cell level. Nucleic Acids Research 52:D107–D114 (2024). doi.org/10.1093/nar/gkad1021

Europe PMC text mining ↑

europepmc · EMBL-EBI · weight ×0.2 · 2,451,412 links · 24,113 genes · 12,259 diseases · homepage

Natural-language processing over millions of abstracts and full-text articles, recording where a gene name and a disease name appear close together, weighted by how unusual the pairing is.

Limitations: co-occurrence is not assertion. The sentence may have said the two are unrelated. It is by far the most abundant evidence type here and carries the lowest weight for exactly that reason — and because it inherits every bias in what people choose to publish about.

• Ferreira R, et al. Europe PMC in 2023. Nucleic Acids Research 51:D1552–D1560 (2023). doi.org/10.1093/nar/gkac1062

Limitations of the annotation layer

Annotation follows attention. A gene studied for forty years carries dozens of annotations; a gene nobody has worked on carries whatever a pipeline could infer from its sequence. The absence of an annotation is not a finding — it is very often a statement about funding rather than about biology.

Generic terms dominate by construction. protein binding is annotated to 13,295 of 20,097 protein-coding genes. BioXplore scores every term by information content — the negative log of the fraction of genes carrying it — and applies a floor, so terms true of everybody drop out without a hand-maintained blocklist.

Annotations inherit upward, and the published inheritance is wrong. Under GO's true-path rule a gene annotated to a term is implicitly annotated to every ancestor of that term. Open Targets publishes an ancestor list that is not the is_a/part_of closure: every apoptosis term lists bleb assembly among its ancestors, which inflates that term from 7 annotated genes to 1,541 and turns a highly specific process into one that scores as generic. 13,370 of 48,321 terms are affected. BioXplore recomputes the closure itself and keeps the published version alongside so the difference stays inspectable.

A process is not a mechanism. That a disease's genes are enriched for regulation of cytokine production says those genes were annotated to that process — not that the process is disrupted in patients, and not in which direction.

• The Gene Ontology Consortium. The Gene Ontology knowledgebase in 2023. Genetics 224:iyad031 (2023). doi.org/10.1093/genetics/iyad031
• Ashburner M, et al. Gene Ontology: tool for the unification of biology. Nature Genetics 25:25–29 (2000). doi.org/10.1038/75556
• Balakrishnan R, et al. A guide to best practices for Gene Ontology curation. Database 2013:bat054 (2013). doi.org/10.1093/database/bat054
• Gaudet P, et al. Phylogenetic-based propagation of functional annotations within the Gene Ontology consortium. Briefings in Bioinformatics 12:449–462 (2011). doi.org/10.1093/bib/bbr042

Limitations of the drug layer

Most genes are not druggable. Only 1,547 of 18,902 disease-associated genes carry any drug at all. An empty drug column is the common case, not a gap in the data.

"Indicated" is not "works". An indication link records that a drug has been approved for or trialled against a disease — including trials that failed. Phase is shown for exactly this reason.

The mechanism is attributed to the annotated target. Many drugs act through additional or unknown routes; the record names the target somebody established, which is not necessarily the only one.

It overlaps the evidence. This layer is not the same thing as the clinical precedence evidence source, but it is built on the same underlying records — so when drug evidence is switched on, the genes in a view were partly selected because drugs target them, and the drug column then shows you those drugs. The explorer warns when this happens and offers to switch the evidence off. Doing so is what turns the column from a restatement into a question.

• Zdrazil B, et al. The ChEMBL Database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods. Nucleic Acids Research 52:D1180–D1192 (2024). doi.org/10.1093/nar/gkad1004
• Ochoa D, et al. Open Targets Platform: supporting systematic drug–target identification and prioritisation. Nucleic Acids Research 49:D1302–D1310 (2021). doi.org/10.1093/nar/gkaa1027

EFO and MONDO

Every disease in BioXplore is an EFO identifier, and most resolve to MONDO — a merged disease ontology reconciling OMIM, Orphanet, DOID and others, so that one disorder has one identifier no matter which resource described it. The hierarchy is what lets the tool exclude a disease's own relatives before ranking its neighbours. A caution that shapes several queries here: the EFO index also contains measurements, phenotypes and biological attributes, which are not diseases and are filtered out explicitly.

• Malone J, et al. Modeling sample variables with an Experimental Factor Ontology. Bioinformatics 26:1112–1118 (2010). doi.org/10.1093/bioinformatics/btq099
• Vasilevsky NA, et al. Mondo: unifying diseases for the world, by the world. medRxiv (2022). doi.org/10.1101/2022.04.13.22273750

Ensembl

The gene models: stable identifiers, symbols, descriptions, genomic coordinates and biotypes. Gene symbols change over time and are not unique across species, so identifiers rather than symbols are the join key everywhere in the catalog; the symbol is a label for display only.

• Harrison PW, et al. Ensembl 2024. Nucleic Acids Research 52:D891–D899 (2024). doi.org/10.1093/nar/gkad1049