Welcome!

Overview

To support target prioritisation, the Open Targets Platform includes tractability data that identifies key details, including if there is a binding site suitable for small molecule binding, an accessible epitope for antibody based therapy, relevant data for using Proteolysis Targeting Chimeras (PROTACs), or a compound in clinical trials with a modality other than small molecule or antibody.

The tractability data can assist in target prioritisation by identifying potential drug targets suitable for discovery pipelines and therapeutic modalities that are most likely to succeed. It also supports further investigation of targets for which there are no ligands or experimental structures or those targets outside a "druggable" target family but with strong genetic associations.

Assessments

Data sources used in the pipeline include UniProt, HPA, PDBe, DrugEBIlity, ChEMBL, Pfam, InterPro, Complex Portal, DrugBank, Gene Ontology, and BioModels.

Assessments common to all modalities, ingested from ChEMBL, are:

• Approved Drug: the target has clinical precedence with Phase IV drugs

• Advanced Clinical: the target has clinical precedence with Phase II or III drugs

• Phase 1 Clinical: the target has clinical precedence with Phase I drugs.

We also include additional assessments specific to each modality.

Small molecule

Antibody

PROTAC

Computational pipeline and datasets

Publications

Overview

Throughout the drug discovery and development process, target safety assessments help with understanding the role of the drug target in normal physiology and potential unintended adverse consequences and safety liabilities when modulating the target with a chemical compound or drug (Brennan, 2017).

Safety data is available on the target profile page and can be accessed to provide a systematic view of potentially relevant target safety liabilities.

Computational pipelines and datasets

Publications

Overview

To support further assessments about the suitability of targets for a discovery pipeline, the Open Targets Platform integrates information from various sources on chemical probes and Target Enabling Packages (TEPs).

Chemical probes

A chemical probe is a small molecule that can act as chemical modulator of a system, such as a cell or an organism, by reversibly binding to a biological target. Chemical probes are expected to have a minimum standard of high affinity, binding selectivity and efficacy. Chemical probes do not need to meet the same requirements in terms of pharmacokinetics, pharmacodynamics, and bioavailability as drugs.

Target Enabling Packages

Target Enabling Packages (TEPs) are a critical mass of reagents and knowledge allowing for the rapid biochemical and chemical exploration of a given target.

• Protein production methods

• Biochemical/biophysical assays for activity, affinity

• Structures of the protein, potentially including wild type and disease mutant proteins; full-length or domains; protein-ligand complexes; structures of close homologues

• Initial chemical matter from a fragment or small molecule screen

Additional components of TEPs may be developed on a case-by-case basis, based upon reasonable scientific need, in collaboration with TEP target nominators:

• An antibody or nanobody

• Cell-based assay

• CRISPR knockout

Data sources

Probes & Drugs Portal

Structural Genomics Consortium

Computational pipelines and datasets

Identifying evidence implicating drug targets with diseases or phenotypes constitutes one of the pivotal challenges of the Open Targets Platform. Several steps are required to aggregate, integrate, validate, and score the collected information, in order to contextualise and weight the underlying evidence. The Platform provides a framework to allow the interactive or programmatic interrogation of target-disease evidence.

Data model

The Open Targets data model focuses on five main entities:

Entity annotation

These entities are annotated with a variety of public data sources, as well as information about the relationships between them. A more detailed description of the annotations is provided throughout the documentation. Some examples are:

Evidence generation and association scoring

Applications and data access

To start addressing therapeutic hypotheses users can find alternative ways to interface with the data:

Training materials

Online tutorial

Please note we are in the process of updating our training content following the most recent updates (25.03).

Webinar

Overview

Pharmacogenetics is the study of how genetic variation may change your response to a specific drug. Through the integration of pharmacogenetics data into the Platform, our objective is to apply clinical annotations available in the Pharmacogenetics Knowledgebase (PharmGKB) to aid target prioritisation for drug discovery.

We also enhance these annotations by adding detailed annotations on variant consequence prediction, drug response categories, and specific drug information and whether the gene is a direct target of the drug. This process involves applying advanced phenotype extraction techniques to offer a refined representation of phenotypes, thereby providing a more precise understanding of the genetic determinants influencing treatment outcomes.

Data sources

We have also included annotations for star (*) alleles, a nomenclature used in the pharmacogenetics field for representing key functional variants involved in drug responses.

Publications

When using this data please remember to acknowledge the sources:

Whirl-Carrillo M, Huddart R, Gong L, Sangkuhl K, Thorn CF, Whaley R, Klein TE. An Evidence-Based Framework for Evaluating Pharmacogenomics Knowledge for Personalized Medicine. Clin Pharmacol Ther. 2021 Sep;110(3):563-572. doi: 10.1002/cpt.2350. Epub 2021 Jul 22. PMID: 34216021; PMCID: PMC8457105.

Whirl-Carrillo M, McDonagh EM, Hebert JM, Gong L, Sangkuhl K, Thorn CF, Altman RB, Klein TE. Pharmacogenomics knowledge for personalized medicine. Clin Pharmacol Ther. 2012 Oct;92(4):414-7. doi: 10.1038/clpt.2012.96. PMID: 22992668; PMCID: PMC3660037.

Overview

Baseline RNA and protein expression data helps us to ascertain whether the target is expressed in all tissues or selectively in one or few tissues or cell types. The availability of the target molecule in the location of interest is critical at different stages of the drug development process.

We combine baseline expression information from three sources:

• Expression Atlas: RNA expression meta-analysis from RNA- sequencing experiments

• Human Protein Atlas: immunohistochemistry-based proteomics data for normal tissues

• Genotype-Tissue Expression (GTEx) Program: RNA baseline expression variation data per tissue

Data sources

Expression Atlas

Human Protein Atlas

Genotype-Tissue Expression (GTEx) Portal

Computational pipelines and datasets

RNA expression meta-analysis

Open Targets in collaboration with Expression Atlas has performed a meta-analysis of over 18,000 samples from 50 different tissues and more than 30 cell types from the following experiments:

The tissue- and cell-based samples in these experiments are processed separately to avoid batch effects during normalisation. The samples of each group are then processed together to generate an expression table of normalised Transcripts Per Million (TPMs) units for every gene in each tissue or cell type according to the following steps:

• Technical replicates are aggregated

• Genes that are expressed below a pre-defined threshold (i.e. minimum of 10 raw reads in at least 15 samples) are filtered out

• • The Coefficient of Variation (CV) was estimated for each gene across all the samples and used to select the least variable genes.

  • The least variable 1,000 genes were used as negative controls; that is, assumed not to be differentially expressed, to train RUVg to remove unwanted variation

• Samples from the same tissue across different experiments are averaged by median and then merged in the final matrix

• Expression tables of the tissue- and cell-based experiments are combined

We analyse this expression file further to compute two values for each gene:

• Binned value of expression: The normalised expression values are divided into 10 bins of the same width. Note that this is not the same as the deciles, which all contain the same number of items in them

• Tissue specificity: Z-scores are calculated for each gene and each tissue and then they are binned based on quantiles of a perfect normal distribution. We compute the tissue specificity of a target as the number of standard deviations from the mean of the log RNA expression of the target across the available tissues. A target is considered to be tissue specific if the z-score is greater than 0.674 (or the 75th percentile of a perfect normal distribution). We remove data for under-expressed targets before the z-score calculation. This allows us to extract the tissues for which a gene is specific, defined as the expression value being above the 75th z-score percentile — in practice, anything in bin 2 or above.

Publications

Overview

Core essential target genes are unlikely to tolerate inhibition and are susceptible to causing adverse events if modulated. This is crucial safety information for drug discovery scientists looking to develop inhibition strategies.

Together with a DepMap target annotation widget, essential target genes are annotated with a `Core essential gene` chip on the target page.

Data source

DepMap Portal

Publications

When using this data please remember to acknowledge the sources:

Pacini C, Dempster JM, Boyle I, Gonçalves E, Najgebauer H, Karakoc E, van der Meer D, Barthorpe A, Lightfoot H, Jaaks P, McFarland JM, Garnett MJ, Tsherniak A, Iorio F. Integrated cross-study datasets of genetic dependencies in cancer. Nat Commun. 2021 Mar 12;12(1):1661. doi: 10.1038/s41467-021-21898-7. PMID: 33712601; PMCID: PMC7955067.

Overview

The Molecular Interactions data aggregates and integrates interaction evidence reported in several resources to provide a systematic view on potentially relevant drug targets. Each of the integrated resources captures relationships of different nature including physical binary interactions, enzymatic reactions, or functional relationships. The information here available aims to capture not only the topology of the interaction network, but also the supporting experimental evidence reported on each of the databases.

In order to maximise coverage, the network contains all reported binary relationships between gene products (proteins and RNAs). Although the main focus are interactions between human molecules, the data also includes additional interactions between human gene products and molecules encoded in the genome of infectious pathogens (viruses and bacteria).

Data sources

IntAct

Interactions are grouped by interaction detection method and interaction type. As a consequence, the same pair of interactors might be split into multiple entries if individual proteins are reported to have different biological roles.

For IntAct, please note:

• The network only contains human and selected pathogen data from IntAct

• The majority of interactions are not directional and not signed. However, there are a proportion of interactions where the biological role of the participants can be stated and directionality specified (e.g. enzymatic reactions)

Reactome

For Reactome, please note:

• Only human-human interactions are provided

• Interactions are directional and signed, with biological roles assigned to each participant if possible

SIGNOR

For SIGNOR, please note:

• SIGNOR only contains human data

• Interactions are directional and signed, with biological roles assigned to each participant

• The network pulls information from the SIGNOR relations file

STRING

Computational pipeline and datasets

The multipartite network displayed in the Open Targets Platform is the result of post-processing the information stored in a Neo4j graph database (graphDB). The graphDB does not provide STRING information but contains ComplexPortal information on stable protein complexes as an additional data source. The information of the graphDB is then exported together with STRINGdb and mapped to the Open Targets Platform targets (Ensembl Gene IDs).

Publications

When using this data please remember to acknowledge the sources:

Orchard S, Ammari M, Aranda B, Breuza L, Briganti L, Broackes-Carter F, Campbell NH, Chavali G, Chen C, del-Toro N, Duesbury M, Dumousseau M, Galeota E, Hinz U, Iannuccelli M, Jagannathan S, Jimenez R, Khadake J, Lagreid A, Licata L, Lovering RC, Meldal B, Melidoni AN, Milagros M, Peluso D, Perfetto L, Porras P, Raghunath A, Ricard-Blum S, Roechert B, Stutz A, Tognolli M, van Roey K, Cesareni G, Hermjakob H. The MIntAct project--IntAct as a common curation platform for 11 molecular interaction databases. Nucleic Acids Res. 2014 Jan;42(Database issue):D358-63. doi: 10.1093/nar/gkt1115. Epub 2013 Nov 13. PMID: 24234451; PMCID: PMC3965093.

Jassal B, Matthews L, Viteri G, Gong C, Lorente P, Fabregat A, Sidiropoulos K, Cook J, Gillespie M, Haw R, Loney F, May B, Milacic M, Rothfels K, Sevilla C, Shamovsky V, Shorser S, Varusai T, Weiser J, Wu G, Stein L, Hermjakob H, D'Eustachio P. The reactome pathway knowledgebase. Nucleic acids research. 2020 Jan;48(D1) D498-D503. doi: 10.1093/nar/gkz1031. PubMed PMID: 31691815. PubMed Central PMCID: PMC7145712.

Licata L, Lo Surdo P, Iannuccelli M, Palma A, Micarelli E, Perfetto L, Peluso D, Calderone A, Castagnoli L, Cesareni G. SIGNOR 2.0, the SIGnaling Network Open Resource 2.0: 2019 update. Nucleic Acids Res. 2020 Jan 8;48(D1):D504-D510. doi: 10.1093/nar/gkz949. PMID: 31665520; PMCID: PMC7145695.

Szklarczyk D, Kirsch R, Koutrouli M, Nastou K, Mehryary F, Hachilif R, Gable AL, Fang T, Doncheva NT, Pyysalo S, Bork P, Jensen LJ, von Mering C. The STRING database in 2023: protein-protein association networks and functional enrichment analyses for any sequenced genome of interest. Nucleic Acids Res. 2023 Jan 6;51(D1):D638-D646. doi: 10.1093/nar/gkac1000. PMID: 36370105; PMCID: PMC9825434.

Overview

Disease or phenotype annotation data sources

Overview

Criteria for target inclusion:

• Genes from all biotypes encoded in canonical chromosomes

• Genes in alternative assemblies encoding for a reviewed protein product.

This definition accounts for some of the complexities of human targets; targets are not only protein coding genes, RNAs or pseudogenes are also considered. However, the current definition has some potential drawbacks. Some drug targets are the result of interactions between genes (e.g. gene fusion) or proteins (e.g. protein complexes). The current target entity does not yet consider these cases, which will be addressed in the future.

Target annotation data sources

Overview

The Clinical Signs and Symptoms data in the Platform aims to capture other diseases or phenotypes that occur as a consequence, or in conjunction with a primary disease.

Disease–phenotype relationships are not only useful to better characterise the disease phenotypic space, but also to serve as proxies for additional causal evidence that can help prioritise new or existing targets.

Data sources

Monarch Merged Disease Ontology (MONDO)

Human Phenotype Ontology (HPO)

Computational pipelines and datasets

Publications

Overview

In the Open Targets Platform, a study is a qualifying genome-wide association study for binary or quantitative traits.

Studies might capture different types of associations, such as curated associations (top hits) from literature or post-processing GWAS summary statistics.

Key study annotation

To provide a rich context to downstream interpretation of studies, and their associations, we capture a wide range annotations, whenever possible in a standardised way to enable interoperability.

For example:

• For molecular QTLs, the affected gene/protein is standardised to Ensembl gene identifiers

• Depending on the granularity of the ingested study metadata, a rich cohort information is captured including sample sizes, ancestry composition and a list of named cohorts

• Indicated if association data of a study was ingested as summary statistics, together with the corresponding quality control measurements performed on the summary statistics

• Besides the Pubmed identifier, rich publication information is also captured including first author, publication year for easier data access

Study categories

Integrated studies can be split into two major groups that determine how their associations would be utilised in the target identification process.

1. Genome-wide association studies (GWAS)

A GWAS identifies associations between genetic variants and traits or phenotypes by analysing genome-wide variation data from a population. These studies cover binary or quantitative traits reported in the integrated sources.

2. molQTL studies

• eQTLs impact gene expression

• pQTLs impact protein abundance

• sQTLs have an effect on gene splicing

• tuQTLs impact transcript usage

• sceQTL impact gene expression at the single-cell level

Each molQTL study captures the unique combination of the following annotations:

• The publication authoring the study

• The gene product measured in the study

• The quantitative method (e.g. aptamer) used to measure the trait, when available

• The cell type or tissue where the trait is measured, when available

• The experimental conditions (e.g. interferon-stimulated macrophages)

Study inclusion criteria

Studies must meet predefined criteria to ensure consistent representation, and increased reliability of associations for downstream analysis.

Each integrated study needs to meet the following criteria:

• GWAS studies need to have a valid disease or phenotype identifier (EFO)

• QTL studies must have an affected gene and tissue/cell-type valid identifier (biosample)

• Valid study type (e.g gwas or a QTL type from a defined set)

• Data licensed for commercial usage

• • Mean beta within expected range

  • Genomic control (GC) lambda value within the expected range

  • The PZ value within the expected range

GWAS/molQTL credible sets

The 'Clinical Precedence' widget in the Open Targets Platform provides users with an overview of the clinical development stage of each drug, as defined by the EMBL-EBI ChEMBL database. This widget is designed to aid in the prioritisation of drug targets by highlighting the extent of clinical development for each drug.

The clinical development stages are categorised as follows:

In the Platform, this information offers valuable insights into the clinical development of potential drug targets. By showing the current phase of clinical development, the 'Clinical Precedence' widget can inform decisions about target prioritisation and help identify drugs that may be suitable for repurposing or further development.

Overview

To further refine how we define a drug in the Platform, we subset ChEMBL's drugbase to capture molecules that meet the following criteria:

• Drugs for which the indication is known

• Drugs for which the target they modulate is identified

• Drugs that are listed in DrugBank

• Drugs that are acknowledged as chemical probes

A drug in the Platform might belong to different modalities, including small molecules, antibodies, or oligonucleotides among others. However, some biologic therapies such as vaccines, blood products, or cell therapies are not represented in our drug set. Moreover, the molecule-centric definition implies multi-ingredient drugs won't be represented and only their individual active moieties might be available on the site.

In the ChEMBL representation of drugs, a clear distinction is made between parent bioactive molecules and their corresponding child molecules. Parent molecules encompass the original, unmodified form of the active ingredient, while child molecules refer to modified versions, such as salts. In the Platform, both parent and child molecules are included, ensuring comprehensive coverage of the drug landscape.

When it comes to data propagation, parent molecules retain their own distinct information, as well as aggregate the specific details from all their child molecules. On the other hand, child molecules solely capture their own individual information, without incorporating any data from other molecules. This consistent approach extends to various aspects, including indications, mechanism of action, and drug warnings. By adopting this systematic framework, our Platform facilitates accurate representation and analysis of the diverse molecular entities and their associated properties.

Drug annotation data sources

Overview

Population Allele Frequencies

Variant effect

Variants are annotated with an integrated view of variant effects from multiple methods. Based on all predictions or annotations, we normalise the variant's likely deleteriousness to a common scale.

To make the predicted variant effects comparable across different methods, raw predictions from each methods were normalised to a unified scale ranging from likely benign to uncertain to likely deleterious.

Transcript consequences

Every variant is annotated with the predicted consequence for all canonical transcripts in a +/-500Kb window, allowing us to understand the likely effects in the neighbouring coding or non-coding genes. For all variant-transcript pairs in the region, this information includes:

• Distance from transcription start site (TSS)

• Distance to footprint

• Predicted functional consequence based on Ensembl VEP

• Amino-acid consequence relative to the UniProt reference protein

Variant-to-phenotype

The list of variant sources includes:

Overview

Before approval, new therapeutic drug treatments are extensively tested in clinical trials. However, some of the side effects are only identified when prescribed to larger cohorts of patients, with one or more medical conditions, for a sustained period of time or in combination with other treatments.

For this reason, regulatory agencies—for example, the Food & Drug Administration or the European Medicines Agency—provide pharmacovigilance programs to monitor and survey Adverse Drug Reactions (ADRs).

Data sources

FDA Adverse Event Reporting System (FAERS)

Computational pipelines and datasets

First we apply a set of filters to the reports as described below:

• Only reports submitted by health professionals (primarysource.qualification in (1,2,3))

• Exclude reports that resulted in death (no entries with seriousnessdeath=1)

• Only drugs that were considered by the reporter to be the cause of the event (drugcharacterization=1)

Next, we sought to map the drugs in the FAERS reports to the drugs in the Open Targets Platform (ChEMBL IDs). Any of the above listed fields were used when exact matches were available:

Due to the nature of the surveillance reports, it's relatively common for the indication for which a drug was prescribed to appear in the list of significant ADRs. Given the current structure of the data provided in a FAERS report, we cannot distinguish whether it's a problem with the dosage the drug was prescribed or an excessive phenotypic characterisation of the patient in the report.

Publications

Each unique target–disease pair in the Open Targets Platform is defined as an association. For example, while there might be several pieces of evidence referring to CFTR and Cystic fibrosis from multiple sources, one single association contextualises all this information within the Platform. Also, since multiple pieces of evidence might refer to the same or similar associations, the Platform undertakes a series of steps to quantify their relative strength for a given association.

Ontological selection of evidence

The Platform associations aim to aggregate all evidence referring to the target and disease, but the complex phenotypic representation of the disease might sometimes cause different pieces of evidence to be annotated against slightly different levels of granularity of the disease.

For example, the association between inflammatory bowel disease and NOD2 is annotated with multiple pieces of evidence referring to these two terms specifically. The Platform refers to associations described by aggregated evidence between two specific terms in our data sources as direct associations. By default, direct associations are displayed in the web application when listing associated diseases or phenotypes with a target of interest.

However, evidence can sometimes be informative to discriminate targets in similar diseases or phenotypes. For example, when evaluating the inflammatory bowel disease and NOD2 association, other pieces of evidence describing the relationship between Crohn's disease and NOD2 might also be informative.

To approach this problem systematically, the Platform makes use of the properties of the disease ontology (EFO), to select all evidence referring to NOD2 in the context of inflammatory bowel disease or any of its ontology descendants (including Crohn's disease). This type of association is referred to in the platform as an indirect association. Indirect associations are displayed in the web application when displaying all the evidence for a target-disease association or listing associated targets with a disease or phenotype of interest.

The same calculations are applied to calculate the association scores, the only difference is the evidence included in the calculation.

Association scores

Deciding what constitutes a strong association is open to interpretation. While some data sources can be more deterministic about the underlying causal evidence, others can occasionally point to targets indirectly linked to the disease. The Platform relies on a series of heuristics to maximise the transparency of the target and disease rankings.

By data source

The Platform's scoring by data source aims to take into account variations in how the data sources organise their evidence.

For example, some data sources are more stringent when it comes to defining what constitutes a single piece of evidence, whereas others rely on their internal evidence score to stratify the strength of the evidence they present.

Additionally, some data sources will capture the meaningful association in one single piece of evidence. In other data sources, the repetition of evidence increases our confidence in the association.

The scoring by data source aims to balance all these differences and provide a consensus view on the strength of the underlying evidence for a particular source.

1. The pieces of evidence are sorted in descending order and assigned an incremental value that indicates their position in the sorted list (the top-scoring item has a positional id of 1, the second has a positional id of 2, and so on).

2. The harmonic sum for each data source is then calculated by summing the result of dividing each evidence score by (positional id^2).

1. To ensure the result is between 0 and 1, the harmonic sum is normalised by dividing the result by the maximum theoretical harmonic sum, which is the one calculated using an infinite vector of ones. The platform derives this calculation (which approximates to 1.644) by using a vector of 1,000 ones.

evidence with score=1.0 -> positional id=1
evidence with score=0.9 -> positional id=2
evidence with score=0.8 -> positional id=3

harmonic sum score = 1.0/1^2 + 0.9/2^2 + 0.8/3^2

max. theoretical harmonic sum score = 1.0/1^2 + 1.0/2^2 + 1.0/3^2 + 1.0/4^2 + ... ~ 1.644
normalised harmonic sum score = harmonic sum score / max. theoretical harmonic sum score

By data type

Overall

Data source weights

Interpreting association scores

There are a few important considerations regarding association scores. As described above, association scores are a heuristic based on the availability of data. While scores are useful to rank lists of targets or diseases, they should not be interpreted as a confidence score for the target-disease association.

For example, under-studied diseases are unlikely to produce high-scoring targets due to the lack of available evidence. In such diseases, a relatively low-scoring target might still be the top-ranked target and potentially a very interesting lead from a therapeutic standpoint.

Similarly, not all associations with available target–disease evidence should be considered legitimate target–disease associations. Some of our data sources rely on predictions to assess the relationship between a target and a disease. Thus, they should be considered with caution and always take their relative support into consideration.

Overview

Pharmacogenetics is the study of how genetic variation may change your response to a specific drug. Through the integration of pharmacogenetics data into the Platform, our objective is to apply clinical annotations available in the Pharmacogenetics Knowledgebase (PharmGKB) to aid target prioritisation for drug discovery.

We also enhance these annotations by adding detailed annotations on variant consequence prediction, drug response categories, and specific drug information and whether the gene is a direct target of the drug. This process involves applying advanced phenotype extraction techniques to offer a refined representation of phenotypes, thereby providing a more precise understanding of the genetic determinants influencing treatment outcomes.

Data sources

We have also included annotations for star (*) alleles, a nomenclature used in the pharmacogenetics field for representing key functional variants involved in drug responses.

Publications

When using this data please remember to acknowledge the sources:

Whirl-Carrillo M, Huddart R, Gong L, Sangkuhl K, Thorn CF, Whaley R, Klein TE. An Evidence-Based Framework for Evaluating Pharmacogenomics Knowledge for Personalized Medicine. Clin Pharmacol Ther. 2021 Sep;110(3):563-572. doi: 10.1002/cpt.2350. Epub 2021 Jul 22. PMID: 34216021; PMCID: PMC8457105.

Whirl-Carrillo M, McDonagh EM, Hebert JM, Gong L, Sangkuhl K, Thorn CF, Altman RB, Klein TE. Pharmacogenomics knowledge for personalized medicine. Clin Pharmacol Ther. 2012 Oct;92(4):414-7. doi: 10.1038/clpt.2012.96. PMID: 22992668; PMCID: PMC3660037.

A more detailed view on the data and methodology is available in:

Overview

A credible set is a set of genetic variants near a genetic association signal that is predicted, with a specific probability, to include the causal variant for that signal. The results of the fine-mapping analysis determine this, assigning each variant in the region a posterior probability of being causal when considering the observed statistics and the population structure. The variants covering the top 95% likelihood of containing the causal variant define the credible sets in the Platform.

A credible set results from statistical analysis on a specific locus in a study. As a consequence, all credible sets are defined as:

• Study in which the association is reported

• Lead variant - Variant with the highest posterior probability in the credible set

• Fine-mapping method and statistics

Credible set exclusion criteria

Credible sets fulfilling any of the next rules are excluded from the Platform:

1. The lead variant is within the MHC region (chr6:25726063-33400556)

3. There is another fine-mapped SuSiE credible set from the same region and study

4. Being a GWAS Catalog fine-mapped top-hit, there is a GWAS Catalog fine-mapped credible set from summary statistics for the same region and study

5. The lead variant is reported in an invalid chromosome (1:22, X, Y, XY, MT)

6. The sum of PIPs in the credible is not within the [0.95,1] range

Credible set confidence

• SuSiE or SuSiE-inf fine-mapped credible set with in-sample LD

• SuSiE or SuSiE-inf fine-mapped credible set with out-of-sample LD

• PICS fine-mapped credible set extracted from summary statistics

• PICS fine-mapped credible set based on reported curated association

• Unknown confidence

The confidence values are symbolised by 0 to 4 stars on the UI. It is important to note that this confidence does not reflect the strength of the association or the effect size.

Credible set variants

All variants in the credible set are annotated with:

• P-value. Unconditioned p-value from the study when available.

• Beta. Corresponds to SuSiE mu (SuSiE fine-mapped GWAS Catalog, UKBB-PPP and FinnGen) or beta (PICS fine-mapped GWAS Catalog and SuSiE fine-mapped eQTL Catalog)

• Standard error. Only available for PICS-fine-mapped credible sets.

• LD (r^2). Linkage-disequilibrium information. Only available for PICS-fine-mapped credible sets.

• Posterior probability. Posterior inclusion probability (PIP) of variant being causal after fine-mapping.

• log(BF). The logarithm of the Bayes factor. Only available for SuSiE credible sets.

• Predicted consequence. The most severe consequence is across all overlapping canonical transcripts, as reported by the Ensembl VEP.

Locus-to-Gene (L2G)

Explaining L2G predictions

We have aggregated the Shapley values into the main feature groups to understand their relative importance.

The base value represents the baseline before any feature-specific information is considered, and is therefore equivalent for all genes and credible sets.

This approach helps to identify which types of evidence (e.g. distance, colocalisation or functional impact) are most influential for a given locus-gene association. In addition, we can visualise the individual contribution of each feature within the group. As the features within the group are highly correlated, the individual values are less interpretable than the group contribution.

A full description of all features is available here.

Colocalisation

Credible sets are compared against other credible sets to find overlapping signals. Two overlapping credible sets are those that share at least one variant in the set. The Platform contains all the overlaps between all GWAS vs all GWAS and all GWAS vs all molQTL studies.

For the overlapping pairs of credible sets, estimates for two co-localisation methods are computed based on the type of credible set:

Colocalisation directionality

A directionality assessment is included in the colocalisation analysis to help the user interpret the relationship between the two overlapping credible sets.

The Platform ingests GWAS, functional genomics and additional datasets to aid the interpretation of the observed signals. These include:

NHGRI-EBI GWAS Catalog

The GWAS Catalog information feeds the Open Targets Platform by providing study metadata and GWAS associations from two sources:

GWAS Catalog literature-curated associations (GCCA)

The harmonisation process includes a number of steps that flag any associations with quality concerns. When harmonising GCCA, the following cases were flagged:

• Variant interaction associations

• Variant location is not available in an association source

• Variant location data is inconsistent (chromosome and position values don’t match)

• Lead variant is duplicated for the same study

• The provided variant location doesn’t match any GnomAD variant

• The risk allele couldn't be mapped to the GnomAD reference

• The lead variant had palindromic alleles

• The lead variant p-value was above the genome-wide significant threshold (p-value>1e-8).

Curated associations were converted into the Gentropy StudyLocus format and uniformly flagged with Study locus from curated top hit quality control flag. Corresponding StudyIndex are generated to captures the study metadata.

GWAS Catalog Summary Statistics (GCSS)

1. filtering out SNPs with unavailable beta, standard error, or p-value

2. filtering out SNPs with zero or Inf values for beta and standard error

3. filtering out SNPs with negative p-values or standard error

4. filtering out SNPs with p-values equal to 1

In some cases, this results in empty SummaryStatisticsdatasets and these studies are excluded from further analysis.

GCSS studies manual curation

The Open Targets team performs additional manual curation for the GCSS studies from “All studies - With study accession numbers, ontology annotations, genotyping technology, cohort identifiers and full summary statistics availability”. The following information is extracted from the corresponding publication:

1. Study type — The GCSS studies identified as pQTL or microbiome GWAS are flagged and excluded from further analysis.

2. Analysis type — studies are flag if performed with any of the following analysis: multivariate analysis, ExWAS, non-additive model, metabolite, GxG, GxE, case-case study. All analysis flags but “metabolite” are excluded from SuSiE fine-mapping and fine-mapped using PICS.

GCSS quality control

1. The P-Z test. This check estimates the mean and standard deviation of the difference between the log p-values reported in the study and the reported betas and standard errors. If at least one value for the study was greater than 0.05, the study fails QC and is flagged with the label The PZ QC check values are not within the expected range.

2. The mean beta check. This check estimates the mean value of the beta across all SNPs in the study. If the absolute mean value is more than 0.05, the study fails QC and is flagged with the label The mean beta QC check value is not within the expected range.

4. Number of variants. All summary statistics with fewer than 2,000,000 variants do not fail QC but are flagged with the label The number of SNPs in the study is below the expected threshold. These studies are not eligible for SuSiE fine-mapping.

FinnGen

eQTL Catalogue

All credible sets and study information are reformatted to Gentropy's StudyIndex and StudyLocus datasets. Credible sets with a lead SNP p-value>=1e-3 are marked with a Subsignificant p-value flag.

The Pharma Proteomics UK Biobank Project (PPP-UKBB)

• SNPs with MAF<1e-4 and INFO<0.8 are filtered out

• Align the order of effective and reference alleles with the gnomAD annotation: if the alleles are reversed, the sign of the effect size is changed; if the allele combination do not match the reference, the SNP is filtered out

• Flag QTLs as cis or trans based on a 5Mb window from the affected gene.

gnomAD

Variant annotation

GnomAD (v4) variant annotation is used to provide extra annotation for variants included in the Platform and assist some of our harmonisation pipelines. Among the most relevant annotations included are: population allelic frequencies, in silico predictors and cross-references. All variants are available in GRCh38 coordinates.

LD matrices

Pan-UKBB LD matrices

EMBL-EBI Ensembl

The Platform uses Ensembl as a source of gene, transcript and variant annotations. Affected genes in QTL studies are validated using fixed version of Ensembl and all variants are annotated using Ensembl VEP.

Biosample

The Experimental Factor Ontology, UBERON and Cell Ontology are composed as a meta-ontology to capture every tissue or cell type that could be described in a QTL study.

Gentropy is an open-source Python package to facilitate the interpretation and analysis of GWAS and functional genomic studies for target identification. The Platform leverages Gentropy to perform post-GWAS analysis and derive the evidence and datasets for web portal visualisation.

Gentropy provides a set of data models, data ingestion methods and statistical analysis organised in discrete steps to maximise re-usability. Gentropy is designed with scalability in mind, so it's suitable for small dedicated analysis but also for the high-performing orchestrated tasks required to generate the data-lake that populates the Open Targets Platform.

Based on genetic and functional genomics traits, the Locus-to-Gene (L2G) machine-learning algorithm ranks the most likely causative genes at each GWAS credible set. The likelihood that a gene is causal for a particular GWAS locus is measured by the L2G score, ranging from 0 to 1.

Features

The predictive features used by the L2G algorithm are designed to capture various genetic and genomic contexts that influence the likelihood of a gene being causal at a given GWAS credible set.

Features are divided into four categories:

Training set

The L2G model is trained based on prior knowledge of gene-trait associations collated from different sources, to later bring representative credible sets supporting that association. The next steps describe the methodology to compose the training set:

Effector gene list

The effector gene list (EGL) represents the set of biologically validated gene-trait associations. This list is derived from several sources:

2. Gene-indication pairs for the pharmacological target in all phase III or IV clinical trials according to the latest ChEMBL release.

To ensure the uniqueness of gene-EFO pairs, the combined list was de-duplicated.

Positives

For each gene-trait pair from the effector gene list, positive gene-credible set pairs were extracted using the following criteria:

1. Only protein-coding genes.

2. Removed any pair with a distanceSentinelTSS feature less than 0.1.

3. Only include gene-trait pairs supported by at least two credible sets in different studies sharing the same lead variant. We added this criterion to reduce the chance of false positives among the credible sets.

4. Among all positive credible set-gene pairs, we removed duplications based on functional genomics features.

5. Removed credible sets involved in more than two positives.

Negatives

All other protein-coding genes in the window are classified as negatives as long as they don't have a strong functional interaction (STRINGdb score > 0.8) with any positive genes for the trait.

The L2G model

The Locus-to-Gene model is trained for every data release using the training set and feature matrix following. Similarly to Mountjoy et al., the model is trained based on a gradient-boosting algorithm with the scikit-learn library, and using nested cross-validation and hyperparameter tuning.

L2G predictions

Reference

Overview

The Platform Bibliography tool aims to provide context on the scientific background available in the literature regarding the target, disease or phenotype or drug entities of interest. The Platform aims to provide not only what publications have referenced the entities of interest, but also what other entities frequently occur in the literature in conjunction with the entities of interest. Open Targets has developed in collaboration with Europe PMC a pipeline that tries to maximise the literature information extraction by using a combination of Natural Entity Recognition, ontological normalisation and data analysis.

On our entity annotation pages, users also have the option of filtering the bibliography data to a specific timeframe.

Data sources

Europe PMC

In conjunction with the Europe PMC corpus, all the information required to build the Platform entities is required in order to establish a dictionary of terms and synonyms associated with each entity.

Computational pipeline and datasets

Natural entity recognition (NER)

A machine learning model trained using BioBERT and a combination of curated datasets is applied on the Europe PMC corpus formed by abstracts and full-text articles. The model tags every gene-protein (GP), disease (DS) or drug (DR) entity in every publication. The resulting set provides some metadata on the publications, the sections where the matches are found and the sentences, where more than one type of entity co-occur.

Entity normalisation

In order to ground the tagged text to the Platform entities, a dictionary-based approach is applied to the tagged entities by applying a series of standard Natural Language Processing tools (e.g. stemming, stop-words, etc.). As a result, a fraction of the tagged sentences are grounded to the Platform entities and annotated with their respective identifier.

Literature-based similar entities

In order to define a strategy to find similar entities based on literature, the resulting matched entities are later used to train a Word2Vec model; the hyper-parameter settings used to train the model are taken following the suggestion from the benchmark conducted by Benjamin et al. (2020).

By using this model, the user can query what are other entities similar to the one selected, based on all the corpus of literature. Similarly, several entities can be selected and the product of their vectors can propose additional entities similar to all the selected entities.

The Bibliography section uses this algorithm to navigate the universe of publications. As a user keeps selecting entities, the universe is narrowed to show the intersection of all the publications that mention all of the selected entities.

OpenAI literature summarisation tool

For data features that link to publications, the Platform now provides the option for users to ask for a natural language summary of the target-disease evidence presented in the publication (when the full-text article is available and free to re-use).

Using LangChain, we ask OpenAI’s GPT-4o mini model to summarise relevant portions of the text which we then ask it to summarise with the following prompt: “Can you provide a concise summary about the relationship between [target] and [disease] according to this study?“. The resulting text is presented to the user (see screenshots with example).

We hope this feature will help the user better understand the available bibliography evidence, and it may actually highlight cases in which the publication does not in fact provide evidence for the target-disease relationship in question.

Publications

Fine-mapping is a statistical analysis technique used to pinpoint the specific genetic variant(s) most likely responsible for a trait association identified in a GWAS. The main result is a credible set (CS): the minimal list of variants with assigned posterior probabilities (PIPs) to be causal that form a predefined probability. The Platform uses 95% CS, meaning the CS has a 0.95 probability of containing the causal variant. The expected sum of PIPs for all variants in CSs have to be within the range of [0.95,1].

Clumping

Clumping is a technique for selecting the most significant variants within a region or set of variants in LD, essentially collapsing signals into loci with an increased chance of containing a single causal variant for the phenotype. Three different methods are used for clumping:

Distance-based clumping

The method is based on an iterative procedure in which the variant with the strongest p-value is selected and all other variants within a predefined distance (radius) from this variant are clumped together. The procedure is repeated as long as there is at least one significant variant. The output of the method is a list of lead variants. There are two parameters: the distance to clump and the p-value significance threshold.

LD-based clumping

The method is applied to the results of the distance-based clumping (list of lead variants). It is again based on an iterative procedure in which the variant with the strongest p-value is selected and all other variants with high LD (r2>=0.5) are clumped together.

Locus breaker

The method consists of three steps:

1. In the first step, we perform the regular distance-based clumping.

2. In the second step, we filter the input summary statistics by the baseline p-value (default is 1e-5). We then clump variants that are closer to each other than the cutoff distance (default is 250,000 bp). Next, we filter clumps by having at least one variant with a p-value above the genome-wide significance threshold. At this stage, we define a list of loci consisting of information about the lead variant and the locus boundaries (the leftmost and rightmost variants in the clump). To each of the locus boundaries we subtract/add (to the left/right boundaries respectively) the flanking distance (the default is 100,000 bp) to avoid situations where the locus size consisting of only one lead variant is 0.

3. In the third step we select the loci with the size greater than the specified large locus size threshold (1,500,000 bp by default) and 'break' each of the large loci using the lead variants from the distance-based clump that lie within the boundaries of the large locus. For each of the distance-based lead variants, we assigned the boundaries as +/-half the size of the large locus size threshold. Thus, each large locus is divided by several overlapping loci of the size of the large locus size threshold. The small loci are unaffected by the splitting.

The general procedure results in a list that contains information about the lead variant and the locus boundaries (the most left and right variants in the cluster), and the largest locus size doesn't exceed the large locus size threshold. The boundaries of the locus are then used to define the region for fine-mapping and LD matrix ingestion. The reason for using this procedure is because locus breaker results in smaller locus sizes on average compared to those defined by window-based clumping.

Acknowledgements: We are grateful to our colleagues at Human Technopole, Sodbo Sharapov and Nicola Pirastu, for advice on this method.

PICS fine-mapping

SuSiE-inf fine-mapping

SuSiE-inf has two approaches for updating estimates of the variance components: Method of Moments and Maximum Likelihood Estimator ('MoM' / 'MLE'). The function takes an array of Z-scores and a numpy array matrix of variant LD to perform fine-mapping. There is a boolean option “est_tausq” that enables the estimation of infinitesimal effects. If it is disabled, it performs fine-mapping similar to the original SuSiE method. We use SuSiE-inf only with the combination of pan-UK Biobank LD matrices for three ancestries: NFE, CSA and AFR.

Fine-mapping pipelines

Different pipeline strategies have been defined for different sources based on a combination of the above methodologies:

Clumping and fine-mapping of GCSS and GCCA PICS

Clumping and fine-mapping of GCSS SuSiE

1. Major ancestry for the study is NFE, AFR, CSA or EAS. If the major ancestry was EAS, we used CSA instead

2. The study type is "gwas"

3. The study has no analysis flags except "metabolite"

4. Study has no quality control flags

5. The locus didn't overlap with the MHC region. We didn’t exclude X or Y chromosomes if they were presented in GWAS

6. The number of variants in the locus after overlapping with the LD matrix was in the range [100, 15,000]

UKBB-PPP clumping and fine-mapping

Pairs of credible sets overlapping at least one variant are subject to additional analysis to infer the likelihood of them sharing the same causal variant. The Platform compares all GWAS vs all GWAS credible sets and all GWAS vs all molQTL credible sets. For overlapping credible sets, co-localisation metrics are estimated using the following methods:

COLOC

• H0: No association with either trait

• H1: Association with trait 1, not with trait 2

• H2: Association with trait 2, not with trait 1

• H3: Association with trait 1 and trait 2, two independent SNPs

• H4: Association with trait 1 and trait 2, one shared SNP

To reduce potential false positives in small overlaps, if the size of the overlap is less than 5 SNPs, we check whether the maximum product of the PIPs for at least one of the overlapping SNP pairs is greater than 0.01. Otherwise, the overlap is excluded from the COLOC analysis.

eCAVIAR

Target–disease evidence

Every event or set of events pinpointing a target as a potential causal gene or protein for a disease represents the unit of information, most often referred to as evidence. Within the Open Targets Platform, a series of pipelines ensure information is retrieved from its sources and standardised in a way that can be immediately applied to answer drug development queries.

All evidence is mapped to the reference target entity identifier (Ensembl gene) and disease or phenotype identifier (experimental factor ontology, EFO), as well as other reference controlled vocabularies and ontologies when appropriate. Evidence is also reviewed to minimise the presence of duplicates within the same data source.

Data sources are also grouped into bigger categories abstracting the type of evidence they predominantly capture. In the platform, these categories are usually referred to as data types, as opposed to the individual resource data referred to as data sources.

The Open Targets Platform provides a scoring framework for each data source to contextualise the relative importance of each piece of evidence. This score will be more relevant when understanding the association scoring in later sections.

Evidence data sources

GWAS associations

The GWAS associations data source aggregates target-disease relationships supported by significant genome-wide associations (GWAS) in the context of other functional genomics data.

Datatype: Genetic associations

Gene Burden

Gene burden data comprises gene–phenotype relationships observed in gene-level association tests using rare variant collapsing analyses. The Platform integrates burden tests carried out by several sources:

• REGENERON (Backman et al., 2021), a whole-exome sequencing analysis of individuals from the UK Biobank.

• AstraZeneca PheWAS Portal (Wang et al., 2021), a whole-exome sequencing analysis of individuals from the UK Biobank.

• Genebass (Karczewski et al., 2022): Gene-based Association Summary Statistics (Genebass), a whole-exome sequencing analysis of individuals from the UK Biobank.

• The results of whole-exome and whole-genome sequencing analysis based on the SPARK cohort bring evidence of novel targets implicated in autism spectrum disorder (Zhou et al., 2022).

• The SCHEMA consortium (Singh et al., 2022), a whole-exome sequencing analysis of individuals with schizophrenia.

• The Epi25 collaborative (Epi25 Collaborative, 2019), a whole-exome sequencing analysis of individuals with epilepsy.

• The Autism Sequencing Consortium (Satterstrom et al., 2020), a whole-exome sequencing analysis of individuals with autism spectrum disorder.

• The results of an Open Targets project (Bomba et al., 2022), a whole-exome sequencing analysis of individuals from the INTERVAL cohort testing for associations between rare coding variants and blood metabolites.

• The results of a pan-ancestry whole-exome sequencing analysis identify relevant genes associated with fat distribution (Akbari et al., 2022).

• The results of whole-exome and whole-genome sequencing analysis on Parkinson disease and promoted by the AMP-PD initiative, and other collaborators (Makarious et al., 2022).

• The results of gene-based analyses of rare variants and circulating metabolic biomarkers relevant to cardiovascular disease (Riveros-McKay et al., 2020).

• The results of rare coding variant analyses from whole exome sequencing of Black South African men to identify genes significantly associated with prostate cancer (Soh et al., 2023)

These associations are a result of collapsing rare variants in a gene into a single burden statistic and regress the phenotype on the burden statistic to test for the combined effects of all rare variants in that gene. The different collapsing methods inform about the filters used to select the set of qualifying variants, mostly based on their pathogenicity and frequency in the population.

Datatype: Genetic associations

Evidence scoring: Scaled p-value from 0.25 (p = 1e-7) to 1 (p < 1e-17).

Direction of Effect assessment:

ClinVar

ClinVar is an NIH public archive of reports of the relationships among human variations and phenotypes, with supporting evidence. The ClinVar data source in the Open Targets Platform captures the subset of ClinVar that refers to germline variants (as opposed to somatic variants). Each evidence in the platform aims to capture an individual RCV record in ClinVar.

Information on variants is covered extensively for both single point and structural variants. When available, genomic coordinates are reported with RS numbers, or by following the CHROM_POS_REF_ALT and HGVS notations.

Datatype: Genetic associations

Evidence scoring: ClinVar evidence is scored in a 2-step process. In Step 1, a score is assigned to every piece of evidence based on the clinical significance:

In Step 2, the score is modulated based on the ClinVar review status:

Direction of Effect assessment:

Genomics England (GEL) PanelApp

The Genomics England PanelApp is a knowledge base that combines crowdsourced expertise with curation to provide gene–disease relationships. Virtual gene panels related to human disorders are reviewed by experts within the clinical and scientific community to support the interpretation of genomes within the 100,000 Genomes Project. Within a panel, genes are rated based on the level of evidence supporting the association with the phenotypes identified by the panel. Genes are then classified according to a traffic light system with red/stop, amber/pause, and green/go classifications. To receive a green rating (diagnostic-grade) on a version 1+ panel, the gene requires "evidence from 3 or more unrelated families or from 2-3 unrelated families where there is strong additional functional data" and "genes that do not meet these criteria are rated as Amber (borderline) or Red (low level of evidence)."

Data type: Genetic associations

Evidence scoring: Based on Genomics England gene rating:

Gene2Phenotype

G2P evidence in the Platform is the result of any target-disease curation by any of the expert panels.

Data type: Genetic associations

Evidence scoring:

Direction of Effect assessment:

UniProt literature

The Universal Protein Resource (UniProt) provides a large compendium of sequence and functional information at the protein level. As part of their functional annotation effort, UniProt curators also annotate proteins with publications supporting their involvement on pathogenic processes.

All publications supporting a given target disease relationship are aggregated into one single Platform evidence.

Data type: Genetic associations

Evidence scoring:

UniProt variants

The Universal Protein Resource (UniProt) also curate variants supported by publications that are known to alter protein function on disease. Curated mutations are predominantly protein coding or in regulatory regions clearly associated with the causal protein.

All publications supporting a given variant in connection with a disease constitute individual evidence. All supporting publications are aggregated within the same evidence.

Data type: Genetic associations

Evidence scoring:

Orphanet

Orphanet is an international network that offers a range of resources to improve the understanding of rare disorders of genetic origin. These resources include an inventory of rare disease and gene associations, classification of the gene–disease relationship, information on the kind of mutation, and supporting publication references.

Data type: Genetic associations

Evidence scoring:

Direction of Effect assessment:

ClinGen

The Clinical Genome Resource (ClinGen) Gene–Disease Validity Curation aims to evaluate the strength of evidence supporting or refuting a claim that variation in a particular gene causes a particular disease. ClinGen provides a framework of guidelines to assess clinical validity in a semi-quantitative manner allowing curators to classify the validity of given gene–disease pair.

All gene–disease pairs mapped to EFO constitute individual evidence in the Platform.

Data type: Genetic associations

Evidence scoring:

ChEMBL

EMBL-EBI's ChEMBL is a manually curated database of bioactive molecules with drug-like properties, either approved for marketing by the U.S Food and Drug Administration (FDA), or clinical candidates. ChEMBL also captures information regarding the drug molecule indications, as well as their curated pharmacological target.

In the Platform, ChEMBL evidence represents any target–disease relationship that can be explained by an approved or clinical candidate drug, targeting the gene product and indicated for the disease. Independent studies are treated as individual evidence.

To provide additional context, we integrate a machine learning-based analysis of the reasons why a clinical trial has ended earlier than scheduled. This sorts the stop reasons into a set of 17 classes which include negative, neutral, and positive reasons. This information is available when hovering on the tooltip of the Source column.

The 17 classes are: Another Study, Business or Administrative, Negative, Study Design, Invalid Reason, Ethical Reason, Insufficient Data, Insufficient Enrolment, Study Staff Moved, Endpoint Met, Regulatory, Logistics or Resources, Safety and Side Effects, No Context, Success, Interim Analysis, and Covid 19.

Data type: Drugs

Evidence scoring: ChEMBL evidence is scored in a 2-step process. In Step 1, a score is assigned to every piece of evidence based on the clinical precedence:

In Step 2, for those clinical trials that have stopped early, the score is down-weighted based on the classification of the reason to stop. In this way, less importance is attributed to evidence of studies that have been stopped due to negative outcomes or safety concerns:

Direction of Effect assessment:

Reactome

The Reactome database manually curates and identifies reaction pathways that are affected by a disease. Reactome annotation includes information regarding the causal target–disease link either being a protein coding mutation or an altered expression.

In the Platform, any mutation or altered expression event affecting a different reaction is captured in a different target–disease evidence.

Data type: Pathways & systems biology

Evidence scoring: All manually curated evidence in Reactome has a score of 1.

CRISPR screens

One of the most powerful approaches to uncover gene function is the experimental perturbation of genes followed by the observation of related phenotypes. The perturbation of gene function in human cells has been greatly facilitated by developments in CRISPR technology.

We have linked cell types to diseases, meaning these diseases are often characterised with abnormal phenotypes in these cell types — hence the association. If knocking out a gene causes significant perturbation in the cell type, it might indicate a potential targeting strategy in the disease.

Data Type: Pathways & systems biology

Evidence Scoring: The Platform uses the linearised CRISPRbrain's assessment of statistical significance to assign a score, including hits from both the upper and lower end of the distribution

Project Score

Project Score is a Wellcome Sanger Institute resource that aims to identify dependencies in cancer cell lines to guide precision medicine. The project combines gene fitness effects derived from whole-genome CRISPR-Cas9 synthetic lethality screenings with tractability data, genomic biomarkers and various target annotation enabling a systematic prioritisation of potential targets. The resulting inferences are then mapped from the cancer cell lines in which the experiment is performed to their corresponding tumours.

In the Platform, any Project Score prioritised target with priority score reaching 36.0 is included as independent evidence; however, pan-cancer dependencies are excluded from the integration.

Data type: Pathways & systems biology

Evidence scoring: Project Score priority score divided by 100

SLAPenrich

In the Platform, each pathway significantly enriched in tumour-occurring mutations constitutes an individual piece of evidence.

Data type: Pathways & systems biology

Evidence scoring: Scaled enrichment p-value from 0.5 (p = 1e-4) to 1 (p<1e-14).

Gene signatures

The Platform also provides information about key driver genes for specific diseases that have been curated from Systems Biology analysis. These publications present different disease gene signatures as potential key drivers or key regulators causing disease.

Data type: Pathways & systems biology

Evidence scoring: Scoring depends on whether the original data contains or not a score:

• p-values and rank-based scores are normalised to the 0.5 - 1 range

• If there is no score a fixed value of 0.5 is used

PROGENy

In the Platform, a PROGENy evidence is defined as any significantly regulated sample-level pathway activities inferred from matched normal vs. tumour samples.

Data type: Pathways & systems biology

Evidence scoring: Scaled p-value from 0.5 (p = 1e-4) to 1 (p<1e-14).

Expression Atlas

The EMBL-EBI Expression Atlas provides a differential expression pipeline aiming to identify genes that are differentially expressed in disease vs control samples. Only contrasts from studies with enough replicates and minimum quality criteria are included in the processing.

In a given contrast, to consider a gene significantly regulated in a contrast, all the following rules are required:

• Absolute log2 fold change > 1

• Adjusted p-value <= 0.05

• Maximum significant genes probes per contrast = 1000

In the Platform, each contrast from independent studies capturing differentially regulated genes constitutes independent evidence.

Data type: RNA expression

Evidence scoring: ExpressionAtlas scoring is the result of the product of:

• Scaled p-value from 0 (p = 1) to 1 (p<1e-10)

• Absolute log2 fold change divided by 10

• Percentile rank divided by 100

Cancer Gene Census

In the Platform, CGC evidence is aggregated at the target–disease level to provide a summary of all curated evidence supporting the involvement of a target with a particular cancer type.

Data type: Somatic mutations

IntOGen

In the Platform, independent target–disease evidence are defined as any significant driver gene detected in any individual cohort. Information regarding the individual driver methods is also provided within each evidence.

Data type: Somatic mutations

ClinVar (somatic)

ClinVar is an NIH public archive of reports of the relationships among human variations and phenotypes, with supporting evidence. The ClinVar (somatic) data source in the Open Targets Platform captures the subset of ClinVar that refers to somatic variants (as opposed to germline variants).

Information on variants is covered extensively for both single point and structural variants. When available, genomic coordinates are reported with RS numbers, or by following the CHROM_POS_REF_ALT and HGVS notations.

Each evidence in the Platform aims to capture an individual RCV record in ClinVar.

Datatype: Somatic mutations

Evidence scoring: ClinVar evidence is scored in a 2-step process. In Step 1, a score is assigned to every piece of evidence based on the clinical significance:

In Step 2, scored is modulated based on the ClinVar review status:

Direction of Effect assessment:

Europe PMC

The EMBL-EBI's Europe PMC enables access to a worldwide collection of life science publications and preprints from trusted sources. The Europe PMC data source aims to identify target–disease co-occurrences in the literature and provide an assessment on the confidence of the relationship. This pipeline uses deep-learning based Named Entity Recognition (NER) to identify gene/proteins and diseases when mentioned in the text, to later normalise them to the target or disease/phenotype entities in the Platform. All co-occurrences of both types of entities in the same sentence are considered evidence.

In the Platform, a piece of Europe PMC evidence is the result of aggregating all co-occurrences of the same target and disease within the same publication.

Data type: Text mining

IMPC

The genotype–phenotype associations made available by the International Mouse Phenotypes Consortium (IMPC) are used to identify models of human disease based on phenotypic similarity scores.

The Wellcome Sanger Institute PhenoDigm is an algorithm aimed at capturing the similarity between a knockout mouse and the clinical manifestations (phenotype) of a human disease. The premise is that if a gene knock-out causes an equivalent phenotype in mouse, the human counterpart is likely to be related with the cause of the disease.

It uses a semantic approach to map between clinical features observed in humans and mouse phenotype annotations. The phenotypic effects in mice are then mapped to phenotypes associated with human diseases. The matches are identified and a similarity score between a mouse model and a human disease is computed.

Data type: Animal models

Direction of Effect assessment:

Cancer Biomarkers

One of the aims of the Cancer Genome Interpreter is to identify how variations in the tumour genome may influence its response to anti-cancer therapies. The Cancer Biomarkers database features biomarkers of drug sensitivity, resistance, and toxicity for drugs targeting specific targets in cancer, curated by clinical and scientific experts in precision oncology, and classified by cancer type.

Data type: Pathways & systems biology

Evidence scoring: All manually curated evidence in Cancer Biomarkers has a score of 1