Target Validation Ladder

Target Validation Ladder — Grading the Human-Genetics Prior

Target choice is the highest-leverage decision in drug development: roughly 90% of clinical programs fail, and the largest share of Phase II/III failures is lack of efficacy — the target hypothesis was wrong, not the molecule. Human genetics is the only widely-validated tool that shifts the prior on efficacy before any clinical data exist, because a genetic variant is a natural, lifelong, randomized perturbation of the target in humans. Targets with genetic support have ~2.6x higher relative success across the pipeline (Minikel, Painter, Dong & Nelson, Nature 629:624, 2024), refining the original ~2x estimate (Nelson et al., Nature Genetics 2015).

The headline 2.6x is an average that hides the structure. The multiplier concentrates where two things hold: (1) the causal gene at the locus is unambiguous, and (2) the genetic direction of effect matches the drug's intended direction. Predictive power is flat across genetic effect size, minor allele frequency, and year of discovery (Minikel 2024) — so a clean common-variant signal can be as informative as a rare large-effect one, provided causality is established. A program citing "GWAS support" without naming the causal gene and stating a concordant direction is claiming far less than it sounds.

This skill turns that into a gradeable rubric. It is the earliest-stage validation check in the suite — it asks "does the human evidence say this target matters, in the right direction, with confidence in the gene?" before any clinical asset exists, then hands the graded prior forward to probability-of-success/mechanism-risk-adjuster, which converts it into a phase-weighted PoS multiplier. The #1 failure mode it exists to prevent: crediting the wrong gene at a locus (~90% of GWAS hits are non-coding; the nearest gene is often not causal).

How to Run

Input

Steps

Grading corpus: Grade a candidate by its nearest analog in references/validated-target-library.md — 9+ primary-sourced targets (PCSK9, HMGCR, NPC1L1, ASGR1, Lp(a), TYK2, TL1A, TREM2, LRRK2…) with evidence type, effect direction, drug outcome, and tier, plus the ladder-calibration multipliers (Nelson 2× → Minikel 2.6×, OMIM RS 3.7, Open Targets dose-response) and a genetics-vs-clinical divergence watch.

Step 1 — Pull the genetics-only association and the pre-computed rank

Establish the prior before grading by hand:

• Open Targets Platform (platform.opentargets.org / Open Targets MCP): pull the target-disease association. Use the 2025 "Associations on the Fly" weighting to zero out non-genetic data sources and read a genetics-only score — non-genetic evidence (literature, pathways) inflates the headline and is not what predicts approval.
• Genetic Priority Score (Mount Sinai / Do lab portal): look up the GPS and GPS-with-direction for the gene × indication (19,365 genes × 399 indications). Anchor: the top 0.28% of GPS conferred ~9.9x odds of being a valid indication and were 1.7 / 3.7 / 8.8x more likely to advance Phase I→II / III / IV (Duffy/Do, Nature Genetics 56:51, 2024).

These two reads give a fast prior. The rest of the workflow confirms why the score is what it is — never trust the number without auditing the causal gene and direction.

Step 2 — Place the evidence on the causality ladder (Axis 1)

The ladder is ordered by causal-gene confidence, not statistical significance. Climb to the highest tier the evidence supports:

EVIDENCE LADDER (highest causal confidence → lowest)

Tier A  Mendelian / rare-variant burden     → OMIM, ClinVar, gene-collapsing
        The variant IS the gene; no ambiguity.   (Regeneron/AZ exome, UKB WES
                                                   + FinnGen 744-endpoint meta, n=653,219)
Tier B  Coding GWAS hit OR colocalization    → Open Targets L2G (0-1, 51 features,
        (coloc posterior >0.8 / L2G >0.5)        92-tissue QTL coloc, Shapley
        Pins the causal gene.                     explanations); coloc/SuSiE; GWAS Catalog
Tier C  cis-MR, colocalized, pleiotropy-robust → cis-MR (Nat Commun 2024); mimics
        Direction + causality, method-dependent   pharmacological perturbation
Tier D  Non-coding GWAS, nearest-gene only    → weakest; the wrong-gene failure mode

Two orthogonal amplifiers sit beside the ladder, not on it:

• Allelic series / dose-response (Tier D in the brief's lettering, Axis 3 here): a graded null→partial→GoF set with monotonic effect is the gold standard — it predicts the shape of the efficacy curve. PCSK9 is the canonical case.
• Human knockouts / LoF tolerance (Axis 4 safety): healthy homozygous/compound-het LoF carriers prove full inhibition is tolerated.

Step 3 — Establish direction-of-effect concordance (Axis 2)

The single most underweighted variable. Association says the gene matters; direction says whether to agonize or antagonize.

LoF protective       ⇒ inhibit / degrade   (PCSK9, ANGPTL3, APOC3, LPA)
LoF causes disease   ⇒ activate / replace  (enzyme replacement, agonism)
Direction discordant ⇒ COLLAPSES the genetic premium — hard gate

Confirm direction with: allelic series, LoF-vs-GoF contrast, and Steiger filtering in MR (orients cause→effect). Demand the direction be stated and concordant with the intended mechanism. Use the GPS-with-direction variant where available. Discordant or unknown direction is the second-most-common reason a "genetically supported" target fails.

Step 4 — Run the safety read-across (Axis 4)

Human genetics predicts harm, not just efficacy:

• gnomAD constraint (v4 807,000; v2 141,456): read pLI and LOEUF (LOFTEE-based). Highly constrained genes (pLI1, low LOEUF) flag that strong/complete perturbation may be poorly tolerated — a caution on full inhibition, not a disqualifier (essential genes can be fine inhibitor targets at sub-maximal engagement).
• Human-knockout catalog: ~3,421 genes (18%) have tolerated two-hit pLoF "human-knockout" genotypes in gnomAD (Minikel et al., Nature 581:459, 2020) — positive evidence of dispensability for inhibitor/degrader programs. Absence is ambiguous (lethality vs. rarity), not disqualifying.
• Pleiotropy as a side-effect predictor: a side effect is ~2x more likely when it resembles a trait genetically associated with the target gene (Duffy/Minikel/Nelson, PLOS Genetics 2025). Operationalize by scanning the gene's full phenome (PheWAS / cis-MR across UKB, FinnGen, MVP — >1M combined) for adverse phenotypes before first-in-human.

Step 5 — Grade modality tractability (Axis 5)

A Tier-A genetically validated target can be small-molecule-intractable; the right read is "validated, needs an antibody/oligo," not "fail." Grade per modality using Open Targets tractability buckets (ChEMBL v34 pipeline):

• Small molecule (8 buckets): approved SM drug → Phase 2/3 → Phase 1 → co-crystal with ligand → high-quality ligand (PFI ≤7) → DrugEBIlity pocket ≥0.7 → DrugEBIlity 0–0.7 → druggable-genome family (Finan 2017).
• Antibody: clinical precedence → high-confidence membrane/secreted localization → predicted signal peptide/TM → Human Protein Atlas membrane evidence.
• PROTAC/degrader: clinical precedence → literature → ubiquitination sites → protein half-life → ChEMBL binder ≤10 µM.
• Oligo / gene therapy (escape hatch): the modern route for high-genetic-confidence / low-classical-tractability targets (e.g., LPA siRNA).

Step 6 — Subtract the failure modes, then score

Before assigning the composite, run the failure-mode checklist (each is a documented reason "genetically supported" targets still fail):

FAILURE-MODE AUDIT
[ ] Wrong gene at locus        → require coding variant or coloc/L2G > threshold
[ ] Wrong direction of effect  → Steiger filter, allelic series, explicit direction
[ ] Horizontal pleiotropy (MR) → cis-only instruments, coloc, MR-Egger/CAUSE
[ ] LD / cross-ancestry confound → matched-ancestry coloc
[ ] Tissue-discordant QTLs     → use disease-relevant tissue
[ ] Reverse causation          → trait causes molecular phenotype, not vice versa
[ ] Indication mismatch        → genetic support for trait X, program targets Y
[ ] Mechanism mismatch         → variant models one isoform/domain only

Then score the six axes (full rubric in references/quick-reference.md) and apply the hard gates.

Output

TARGET VALIDATION LADDER — [GENE] in [Indication]
Date: [assessment date]    Intended mechanism: [inhibit/agonize/...]    Modality: [...]

Prior reads:
  Open Targets genetics-only score: [0-1]
  Genetic Priority Score percentile: [%]  (top 0.28% ⇒ ~9.9x indication odds)

Axis scores:
  1. Causal-gene confidence (0-4):        [n]  — [ladder tier + evidence]
  2. Direction concordance (0-2):         [n]  — [LoF protective ⇒ inhibit, etc.]
  3. Allelic series / dose-response (0-2):[n]  — [series / partial / single variant]
  4. Safety read-across (0-2):            [n]  — [pLI/LOEUF; KO tolerance; PheWAS pleiotropy]
  5. Tractability (0-2, per modality):    [n]  — [OT bucket]
  6. Replication / generalizability (0-1):[n]  — [biobanks / ancestries]

  COMPOSITE: [0-13]  →  TIER [1-5]: [de-risked / strong / emerging / hypothesis / none]

Hard gates applied:
  [ ] Discordant direction → capped at Tier 3
  [ ] Non-coding-only, no coloc → causal confidence capped at 1
  [ ] Intractable in all modalities → flag undevelopable

Failure-mode flags: [list any tripped]

Verdict: [PCSK9-class de-risked / strong with gaps / emerging needs orthogonal work /
          association-grade hypothesis / no human prior]

Hand-forward to mechanism-risk-adjuster:
  Genetic-support tier: [Tier 1-5]
  Implied relative-success multiplier: [~2.6x top-bin → ~1.0x for Tier 5]
  Direction-concordant: [Y/N]    On-target safety flag: [clean / caution / adverse]

Error Handling

Cross-Domain Connections

This skill is deliberately dual-use — the same rubric serves two readers:

• A clinical scientist learning the field uses it as a teaching ladder: it makes explicit why PCSK9 (Tier A allelic series + healthy human knockout) is a different class of evidence from a non-coding GWAS hit, and why direction-of-effect is the variable that separates a real agonist/antagonist call from a coin flip.

• An investor screening early opportunities uses it as a diligence gate: a weak rung on this stack is a red flag even when the preclinical data are pretty, and the composite tier becomes the genetic-support input to the conviction stack — applied before a clinical asset exists, when the cheapest de-risking decision is made.

• probability-of-success/mechanism-risk-adjuster (depends_on — primary hand-forward): This skill runs earlier. It grades the human-genetics prior; the risk-adjuster consumes the graded tier and converts it into the PoS multiplier (genetic target validation ≈ +20-30% / ~2.6x relative success). The clean contract: this skill outputs {genetic-support tier, direction-concordant Y/N, safety flag}; the adjuster turns it into a phase-weighted number. Never let the adjuster apply a genetic premium this skill has not graded.

• probability-of-success/pos-base-rates: supplies the therapeutic-area / modality base rate that the 2.6x genetic multiplier scales.

• modality-trajectory/modality-lifecycle (sibling): when a Tier-A target is intractable by small molecule/antibody, the lifecycle map says whether the rescuing modality's delivery unlock has landed (oligo, degrader, gene therapy).

• modality-trajectory/moa-analog-engine (sibling): a genetics-first target (PCSK9/GPR75 pattern) is one of three arc templates; this skill confirms the "genetics-first" classification with graded evidence.

• frontier-discovery/frontier-conviction-scorer: the composite tier is the P(biology holds) term in the discovery-stage trajectory score.

• competitive-intelligence/pipeline-mapper: a high-tier target with thin clinical competition is white space; pipeline crowding around a Tier-A target is mechanism validation.