Drug Target Identification

Human Genetics & Functional Genomics

The most direct and most clinically-validated route: use natural human variation and deliberate genetic perturbation to name a gene and its direction.

The organizing evidence: genetic support ≈ 2–2.6× approval success

• Nelson et al. 2015 (Nat Genet 47:856, 10.1038/ng.3314) — the proportion of drug mechanisms with direct genetic support rises across the pipeline from 2.0% (preclinical) → 8.2% (approved); genetically-supported targets are ~2× more likely to be approved. Method: mapped OMIM + GWAS gene–disease pairs onto drug target–indication pairs by MeSH similarity.
• King, Davis & Degner 2019 (PLoS Genet 15:e1008489, 10.1371/journal.pgen.1008489) — independent replication; confirms the ~2× effect from Phase I to approval.
• Minikel et al. 2024 (Nature 629:624, 10.1038/s41586-024-07316-0, PMID 38632401) — current state of the art: relative success 2.6×; the effect grows with confidence in the causal gene, varies by therapy area/phase, but is largely independent of effect size, allele frequency, and year of discovery. Conclusion: the field is "far from peak genetic insight."

This is why "which gene is causal, and how sure am I" — the output of L2G/coloc/MR/burden below — is the currency of modern target discovery.

GWAS → gene mapping (the core problem)

How the method works

A GWAS returns an associated locus (an LD block of correlated SNPs), not a target. ~90% of trait-associated variants are non-coding, acting through cis-regulatory DNA to change expression, not protein sequence (Maurano et al. 2012, Science 337:1190, 10.1126/science.1222794). Two hard sub-problems: fine-mapping (LD means the lead SNP is rarely causal → statistical fine-mapping with SuSiE/FINEMAP/CAVIAR reduces a locus to a credible set with posterior inclusion probabilities) and the nearest-gene problem (enhancers skip nearby genes, act over >1 Mb, hit multiple genes, are cell-type-specific — distance is a prior, not proof). Closing the gap requires integrating distance + coding consequence (VEP) + molecular-QTL colocalization + chromatin contact + causal-direction (MR). Reviewed in Cano-Gamez & Trynka 2020 (Front Genet 11:424, 10.3389/fgene.2020.00424).

Limits. Causal-variant resolution is LD-bound; regulatory effects are context-specific; QTL catalogs miss the right cell state; no single evidence line is decisive.

Locus-to-Gene (L2G) — ML prioritization

Paper. Mountjoy et al. 2021, Nat Genet 53:1527, 10.1038/s41588-021-00945-5. (Correction: the actual title is "An open approach to systematically prioritize causal variants and genes at all published human GWAS trait-associated loci," not the commonly-miscited "annotating and prioritizing… at scale" string.)

How the method works

A gradient-boosting classifier scores each gene near a fine-mapped credible set for probability of being the causal effector → an L2G score 0–1. Features (weighted by variant posterior probability): TSS distance (absolute + relative to neighbors), molecular-QTL colocalization (eQTL/pQTL/sQTL summarized as CLPP/H4), VEP consequence severity, enhancer–promoter chromatin contact, local gene density, credible-set confidence. Trained on gold-standard positives (curated causal loci + ChEMBL Phase III/IV drug targets + high-confidence ClinVar/UniProt/Gene2Phenotype/PanelApp/ClinGen). Now runs in the open-source Gentropy pipeline with Shapley values exposed per prediction.

Strengths/limits. Recovers held-out ChEMBL Ph III/IV targets; genome-scale, calibrated, open. But only as good as its QTL/chromatin inputs (tissue gaps), trained on known biology (may under-rank novel mechanisms), and distance remains a dominant feature (inherits nearest-gene bias where functional data are sparse).

Colocalization (eQTL/pQTL coloc) and SMR

• coloc — Giambartolomei et al. 2014, PLoS Genet 10:e1004383, 10.1371/journal.pgen.1004383. Bayesian test over 5 hypotheses; PP.H4 = one shared causal variant drives both the GWAS signal and a molecular QTL (the effector-gene nomination); PP.H3 = two distinct LD variants (the false positive it guards against). pQTL coloc is one step closer to the druggable entity than eQTL; cis-pQTLs are strongest.
• SMR + HEIDI — Zhu et al. 2016, Nat Genet 48:481, 10.1038/ng.3538. Treats the top eQTL as an MR instrument for expression→trait; HEIDI separates a single shared causal variant (causality) from linkage. Prioritized e.g. TRAF1/ANKRD55 (RA), SNX19/NMRAL1 (schizophrenia).

Tools/DBs. coloc, SMR/HEIDI; GTEx & eQTLGen (eQTL); deCODE / UKB-Olink / SomaScan (pQTL); embedded as L2G features in Open Targets. Limits: classic coloc assumes one causal variant per locus (fixed by SuSiE-coloc/conditioning); needs large QTL samples in the right tissue; gives a shared variant but not causal direction (MR adds that).

Mendelian Randomization (drug-target / cis-MR)

Framework. Schmidt et al. 2020, Nat Commun 11:3255, 10.1038/s41467-020-16969-0; practical guide Gill, Burgess et al. 2021, Wellcome Open Res 6:16, 10.12688/wellcomeopenres.16544.2.

How the method works

Germline variants randomized at conception act as instrumental variables for an exposure (protein/expression level) — a natural experiment that can reduce reverse causation and some confounding, subject to instrumental-variable assumptions. Drug-target cis-MR instruments the exposure with variants in/near the target's gene, so the genetic perturbation can inform a hypothesis about modulating that protein and the direction + magnitude of the disease effect (potential efficacy and on-target safety) before a trial. Cis restriction anchors the no-horizontal-pleiotropy assumption.

Canonical proofs. PCSK9/HMGCR (cis-LDL-lowering variants → lower CHD, mirroring statins & anti-PCSK9 mAbs; foundation Cohen et al. 2006, NEJM 354:1264, 10.1056/NEJMoa054013 — PCSK9 LoF → ~88% CHD reduction; MR also flagged the on-target T2D liability); IL6R (Asp358Ala mimics tocilizumab; IL6R MR Consortium 2012, Lancet 379:1214, 10.1016/S0140-6736(12)60110-X); IL23R (protective LoF R381Q validated ustekinumab; Di Meglio et al. 2011, PLoS ONE 6:e17160, 10.1371/journal.pone.0017160).

Limits. Three IV assumptions; horizontal pleiotropy is the main threat (cis reduces but doesn't eliminate; run coloc alongside); estimates a lifelong average effect (≠ a drug's dose/timing).

Rare-variant burden / exome association (names the gene directly)

How the method works

Single rare variants are too infrequent for GWAS; gene-based tests collapse all rare variants in a gene into one unit, naming the gene = the target: burden/collapsing (CMC), variance-component SKAT, and the data-driven optimal mixture SKAT-O. LoF variants = "natural knockouts" — a lifelong human analog of pharmacological inhibition; if LoF carriers show a desirable phenotype, inhibition is a hypothesis supported by human genetics: PCSK9 (low LDL, CAD protection), ANGPTL3 (familial hypolipidemia → evinacumab, 10.1056/NEJMoa2004215), SLC30A8 (LoF protects from T2D → direction is inhibition).

Strengths/limits. Names the gene + gives direction + a human tolerability read in one shot; needs 100k+ exomes; the "qualifying variant" definition (LoF-only vs +missense) strongly drives results; loses power under mixed-direction effects.

Mendelian disease genes (OMIM)

How the method works

A gene that causes a monogenic disease proves perturbing it moves human physiology. Match the direction (GoF disease → inhibitor; LoF → activator/replacement; dominant-negative → caution). An allelic series (mild→severe variants → graded phenotypes) is a human dose-response curve. OMIM: Amberger et al. 2019, NAR 10.1093/nar/gky1151; direction-of-effect systematized in npj Drug Discovery 2025 (10.1038/s44386-025-00027-0). Limits: monogenic biology may not translate to common-disease dosing.

The substrate & integration layer (biobanks + platforms)

Biobanks:

Exome/phenome portals. Genebass (Karczewski et al. 2022, Cell Genomics 10.1016/j.xgen.2022.100168, PMID 36778668) — SAIGE-GENE across 4,529 phenotypes × 394,841 UKB exomes; AZ PheWAS Portal (Wang et al. 2021, Nature 10.1038/s41586-021-03855-y) — collapsing analysis, 281,104 exomes, 1,703 gene–phenotype hits (median OR 12.4), 83% invisible to single-variant tests (the headline argument for gene-based tests).

Integration layer. Open Targets Platform aggregates GWAS + rare-variant + Mendelian + functional + literature into scored target–disease associations (Ochoa 2023 NAR 10.1093/nar/gkac1046; 2025 update 10.1093/nar/gkae1064). Key 2024 change: Open Targets Genetics was deprecated and merged into the unified Platform, which now natively carries GWAS credible sets, the L2G model, direction-of-effect, PROTACtability, and AlphaFold structures. GWAS Catalog (Sollis 2023 NAR 10.1093/nar/gkac1010) feeds L2G/coloc; PheWAS inverts GWAS (fix a gene, scan all phenotypes) to reveal pleiotropy = on-target side-effect prediction.

Functional genomics — perturbation screens (deliberate genetic KO/KD/activation)

• Genome-wide CRISPR-KO — pooled sgRNA library into Cas9+ cells; dropout = essential/required, enrichment = resistance. Foundational: Shalem 2014 (Science 343:84, 10.1126/science.1247005, GeCKO); Wang 2014 (Science 343:80, 10.1126/science.1246981); sgRNA design Doench 2016 Rule Set 2 (10.1038/nbt.3437); hit-calling MAGeCK, Li 2014 (10.1186/s13059-014-0554-4). Limit: cutting at high-copy loci causes a DNA-damage copy-number artifact (false essentiality), corrected by CERES/Chronos.
• CRISPRi / CRISPRa — catalytically-dead dCas9; dCas9-KRAB (tunable knockdown, models drug inhibition better than a full null and avoids the cutting artifact), dCas9-activators (uniquely nominate gain-of-function/overexpression targets). Qi 2013 (10.1016/j.cell.2013.02.022); Gilbert 2014 (10.1016/j.cell.2014.09.029); Horlbeck 2016 (eLife, 10.7554/eLife.19760).
• Perturb-seq — CRISPR perturbation with single-cell-transcriptome readout: turns "does KO kill?" into "what pathway is this gene in?" Dixit 2016 (10.1016/j.cell.2016.11.038); Adamson 2016 (10.1016/j.cell.2016.11.048); genome-scale Replogle 2022 (Cell 185:2559, 10.1016/j.cell.2022.05.013, PMID 35688146, >2.5M cells).
• RNAi — the CRISPR predecessor; seed-region off-target silencing (Jackson 2003, 10.1038/nbt831) drove the field to CRISPR after 2014; legacy datasets remain minable.
• In vivo CRISPR screens — capture microenvironment/immune selection; Manguso 2017 (Nature 547:413, 10.1038/nature23270) nominated PTPN2 as a checkpoint-immunotherapy target.
• Base-/prime-editing screens (variant-to-function) — ask "what does this nucleotide do," classifying VUS and saturation-mutagenizing residues. Hanna 2021 (Cell 184:1064, 10.1016/j.cell.2021.01.012); benchmark Findlay 2018 BRCA1 saturation genome editing (Nature 562:217, 10.1038/s41586-018-0461-z).

Genetic-perturbation → target logic (turning screens into nominations)

• DepMap / Project Achilles — genome-scale CRISPR-KO across 1,000+ cancer lines; each gene×line gets a gene-effect score. The nomination logic: pan-essential genes (ribosome/proteasome) are bad targets (toxic, no window); selective dependencies (essential only in a biomarker-defined subset) are the nominations. Defining paper Tsherniak 2017 (Cell 170:564, 10.1016/j.cell.2017.06.010); scoring CERES, Meyers 2017 (10.1038/ng.3984) → Chronos, Dempster 2021 (10.1186/s13059-021-02540-7). Sanger counterpart Project Score, Behan 2019 (Nature 568:511, 10.1038/s41586-019-1103-9). Toxicity logic formalized in Chang et al. 2020 (Cancer Cell 38:171, 10.1016/j.ccell.2020.12.017).
• Synthetic lethality — drug the SL partner of an undruggable tumor loss → selective kill. PARP/BRCA is the validated paradigm: Bryant 2005 (10.1038/nature03443) + Farmer 2005 (10.1038/nature03445) → olaparib FDA-approved Dec 2014 with companion Dx. WRN–MSI: two screens ranked WRN the top dependency in MSI cancers (Chan 2019, Nature 568:551, 10.1038/s41586-019-1102-x; Behan 2019) → WRN inhibitors now clinical.
• Dependency / co-dependency mapping — correlate a gene's dependency profile with genomic features (→ biomarker) or with another gene (→ paralog buffering): MAGOH/MAGOHB (Viswanathan 2018, 10.1038/s41588-018-0155-3), VPS4A/VPS4B (Neggers 2020, 10.1016/j.celrep.2020.108493) — exploiting collateral lethality from passenger deletions.
• Cell-context specificity — oncogene addiction (Weinstein 2002, 10.1126/science.1073096; BCR-ABL→imatinib, BRAF-V600E→vemurafenib — highest-confidence class) and lineage dependency (Garraway & Sellers 2006, 10.1038/nrc1947; SOX10/MITF in melanoma).

Omics-Driven & Network / Systems-Biology Approaches

Nominate candidates from the molecular readouts of disease, then consolidate scattered hits into modules and mechanisms. Strongest when the output is credentialed against genetics (Part A).

Transcriptomics

• Differential expression (DE) — negative-binomial GLMs with empirical-Bayes shrinkage: DESeq2 (10.1186/s13059-014-0550-8), edgeR (10.1093/bioinformatics/btp616), limma-voom (10.1186/gb-2014-15-2-r29). First-pass shortlist; limit: bulk DE averages over cell types, significance ≠ causal ≠ druggable, mRNA ≠ protein.
• WGCNA / co-expression modules — soft-threshold → TOM → modules (eigengenes); targets via module–trait correlation, hub genes (kME), module preservation (Zsummary). Langfelder & Horvath 2008 (10.1186/1471-2105-9-559). Limit: correlation ≠ causation; composition-confounded.
• Connectivity / reverse-signature (CMap & LINCS L1000) — a perturbagen whose signature anti-correlates with the disease signature is predicted to revert it (nominating drug + genetic MOA). Lamb 2006 (Science 10.1126/science.1132939); Subramanian 2017 (Cell 10.1016/j.cell.2017.10.049); query via Broad CLUE. Limit: cell-line ≠ disease tissue; correlational.
• Master-regulator / driver activity (ARACNe · MARINa · VIPER) — infers a regulator's activity from its regulon's collective differential expression, surfacing drivers whose own mRNA is unchanged. ARACNe (10.1186/1471-2105-7-S1-S7); VIPER, Alvarez 2016 (Nat Genet 10.1038/ng.3593). Per-sample (VIPER/OncoTarget/OncoTreat) → patient-level prioritization.

Proteomics & other omics (closer to the druggable layer)

• Mass-spec proteomics (DDA/DIA/TMT) — quantifies the actual druggable molecule; DIA (Orbitrap Astral/timsTOF + DIA-NN) is now the default, near-transcriptome depth. Limit: dynamic range; abundance ≠ activity.
• Phosphoproteomics / kinase-activity inference — substrate-based inference (KSEA; Wiredja 2017 10.1093/bioinformatics/btx415; PhosX 2024) flags hyperactivated/bypass kinases.
• Proteogenomics (CPTAC) — matched genomics + transcriptomics + (phospho)proteomics filters genomic hits for protein-level over-expression/hyperactivation; Savage et al. 2024 (Cell 187:4389, S0092-8674(24)00583-X, PMID 38917788) — the leading rigorous multi-omic nomination framework (beats transcript-only).
• Chemoproteomics / target deconvolution (TPP-CETSA, ABPP) — identify the direct target of a small molecule in native proteomes: TPP (Savitski 2014, Science 346:1255784, 10.1126/science.1255784, PMID 25278616 — staurosporine engaged >50 kinases; ferrochelatase off-target explained vemurafenib phototoxicity); ABPP (Cravatt). The go-to for deconvolving phenotypic-screen hits, covalent ligands, PROTAC/molecular-glue programs.
• Metabolomics — perturbed metabolite pools → responsible enzymes/transporters. Limit: metabolite→enzyme mapping is ambiguous.
• Epigenomics (ATAC/ChIP/methylation; super-enhancers) — nominate epigenetic readers/writers/erasers (BRD4, EZH2) and super-enhancer-addicted oncogenes/master TFs (BET/BRD4 inhibitors).

Single-cell & spatial

• Cell-type-specific target ID (scRNA-seq) — is a gene cell-type-specific in healthy tissue (cleaner target) and over-expressed in a disease cell type; doubles as a safety filter (expression in a vital cell type → predicted on-target toxicity). scRNA-supported genes enrich for clinical success (Open Targets scRNA framework).
• Disease-associated cell states — a pathogenic state's up-regulated surface/secreted/TF genes become candidates; state abundance is a built-in PD marker (DAM → TREM2, now an active CNS target; Keren-Shaul 2017, [10.1016/j.cell.2017.05.018 / S0092-8674(17)30578-0]). Differential abundance via Milo (10.1038/s41587-021-01033-z).
• Cell–cell interaction / ligand–receptor — CellPhoneDB, NicheNet (10.1038/s41592-019-0667-5, links sender ligands to receiver target genes → causal ranking), CellChat. Limit: co-expression ≠ physical binding; needs spatial confirmation.
• Spatial transcriptomics — where a target is expressed, which niches form; validates L-R adjacency. Visium/Slide-seq (coverage) vs MERFISH/Xenium/CosMx (single-cell). Limit: resolution-vs-coverage tradeoff; correlative.

Multi-omics integration

Taxonomy: early (concatenate), intermediate/joint (shared latent space — dominant for target-ID), late. Core methods: MOFA/MOFA+ (Bayesian PCA-generalization; factors → drivers; Argelaguet 2018/2020, 10.15252/msb.20178124, 10.1186/s13059-020-02015-1); SNF (fuse patient-similarity networks; Wang 2014, Nat Methods 10.1038/nmeth.2810); iCluster/iClusterBayes (joint latent-variable with feature selection; Shen 2009, Mo 2018); MOGONET (per-omic GCNs + view-correlation; Wang 2021, 10.1038/s41467-021-23774-w). (These same methods recur in the stratification survey — there they define patient subtypes; here their sparse loadings/selected features name driver targets.)

Network & systems biology

• PPI networks — disease proteins cluster in interactome neighborhoods; rank by degree/betweenness/centrality. STRING ([10.1093/nar/gkac1000-class], 2025 release adds regulation directionality), BioGRID, IntAct, HuRI. Limit: interactome is incomplete/study-biased (well-studied proteins look like hubs).
• Network propagation / RWR — seed signal (GWAS/DE/mutation) diffuses over edges; genes near many seeds score high (HotNet2, DIAMOnD). Unifying review Cowen et al. 2017 (Nat Rev Genet 18:551, 10.1038/nrg.2017.38). Limit: inherits interactome bias; needs degree correction.
• Key Driver Analysis (KDA) — overlay disease genes on data-driven networks; genes whose neighborhood is disease-enriched are candidate master regulators (Mergeomics; Shu 2016 10.1186/s12864-016-3198-9).
• Pathway/GO enrichment (GSEA) — MOA context for candidates (Subramanian 2005, PNAS 10.1073/pnas.0506580102).
• GRN inference (GENIE3 · SCENIC · VIPER) — reconstruct TF→target edges; single-cell SCENIC/VIPER shift target-ID to "which gene, in which cell state."
• Network medicine (Barabási) — disease-module hypothesis + drug-target–module proximity for repurposing/nomination (Barabási 2011, 10.1038/nrg2918; Guney 2016, 10.1038/ncomms10331).

AI / Knowledge-Graph Methods & the Industry Platforms

How companies apply these methods, and how their reported programs differ in clinical maturity.

Method families (how AI/KG nominates targets)

• Biomedical knowledge graphs — ingest genes/proteins/diseases/drugs/trials/literature as typed nodes+edges; query graph patterns + embeddings to surface mechanistic hypotheses. Operationalized by BenevolentAI (below).
• Literature-based discovery / LLMs over text — mine publications/patents/trials for latent gene–disease links and novelty scoring (Insilico's NLP layer; emerging LLM agents).
• Deep target-prediction & structure/druggability — protein language models (BioMap's xTrimoPGLM), AlphaFold-family structure prediction (Isomorphic), cofolding/pose models (Genesis Pearl, Iambic NeuralPLexer) — mostly target-adjacent (how to drug a chosen target), not de-novo target ID.
• Phenomics / perturbation-at-scale — Recursion's image-based maps of biology; Cellarity's cell-state models; hypothesis-free gene–gene/gene–compound nomination.
• Federated multi-omics — Owkin's TargetMATCH across hospital pathology+omics without pooling data.
• Causal ML + functional genomics — Relation's "Lab-in-the-Loop" causal graph models validated by CRISPR.

Industry platforms and clinical maturity (August 2026)

Genuine target-ID engines: Insilico/PandaOmics, BenevolentAI, Recursion, Verge, Owkin, Relation, Cellarity, BioMap (partial). Molecule-design/structure companies that work known targets (not target-ID): Genesis, Iambic, Isomorphic, Nimbus.

Interpreting company programs

1. Insilico's TNIK/rentosertib is the only AI-nominated novel target with a positive peer-reviewed Phase 2 — and even there, FVC efficacy was a secondary endpoint in a small China-only trial (Nat Biotechnol 2024 10.1038/s41587-024-02143-0; Nat Med 2025 10.1038/s41591-025-03743-2, PMID 40461817); Phase 3 initiated Jul 2026.
2. BioMap–BMS deal is unverifiable — the widely-cited "BMS AI-protein deal (~$400M, Dec 2024)" was with AI Proteins, Inc., a different company.
3. "GS-100" is Grace Science's NGLY1 gene therapy, not Genesis Therapeutics' — Genesis has no named clinical asset.
4. Verge's ALS co-development partner is Ferrer (Mar 2024), not Eli Lilly.
5. Cellarity's lead is CLY-124 (DCN1 inhibitor, SCD; Ph1 dosed Jun 2025), not "CLTY-2003."
6. Firsocostat IS the Nimbus/Gilead ACC molecule; Iambic "IAM-H1" is not a distinct program (a HER2 database entry, likely = IAM1363).

Clinical interpretation: Early clinical evidence should be distinguished from confirmatory efficacy and from validation of a target-discovery platform. Insilico's TNIK is the cleanest arc (novel target → +Ph2 → Ph3, pending the real Ph3 test); Nimbus's zasocitinib and Iambic's IAM1363 are the strongest clinical results but validate molecule design on established targets, not target discovery; Owkin's best evidence is a diagnostic; Recursion, despite the largest dataset, has no Phase-2 efficacy win from a self-nominated target. Platform-productivity claims across all companies remain [company-reported] pending mature clinical outcomes.

Building an integrated target-discovery strategy

Across all three parts the same meta-pattern recurs — no single method wins; the strongest programs chain them:

1. Human genetics is the anchor (best single predictor of clinical success; ~2/3 of approvals genetically supported). Name the causal gene, pin the direction.
2. Single-cell / spatial adds "which cell type / state" — now expected, not optional; doubles as a safety filter.
3. Omics (transcript → protein → metabolite → epigenome) nominate candidates; protein/activity-level evidence (proteogenomics, VIPER, phospho, chemoproteomics) outranks transcript-only.
4. Networks add mechanism/context (propagation, KDA, module proximity) and turn scattered hits into modules + repurposing hypotheses.
5. Perturbation (CRISPR screens, Perturb-seq, DepMap, TPP-CETSA) converts markers into tested dependencies and deconvolves direct targets.
6. Direction-of-effect + modality + structure (Open Targets, AlphaFold, PROTACtability) decide what is actually actionable.

Integration platforms — Open Targets above all — rather than any single method are what practitioners actually deploy. AI/KG platforms are a fast-moving frontier that so far augments, not replaces, this genetics-anchored pipeline; the evidence reviewed here is that topology narrows the list, but genetics + structure + modality decide what's actionable — and clinical evidence for each target and molecule must be assessed separately.

Interpreting the evidence

Industry-reported productivity metrics and early clinical signals represent different levels of evidence. A successful molecule against an established target does not, by itself, validate a target-discovery method. Company programs and clinical stages are described as of August 2026.