Impact of proteogenomic evidence on clinical success

This study demonstrates that incorporating plasma protein quantitative trait loci (pQTL) evidence significantly enhances the clinical success rate of therapeutic targets, increasing the likelihood of advancing from Phase I to launch by 4.7-fold compared to the 2.6-fold improvement seen with human genetic evidence alone.

The Gist

Imagine you are a detective trying to solve a massive mystery: Why do some new medicines work, while others fail?

For years, scientists have had a "clue" called Human Genetic Evidence. Think of this like finding a fingerprint at a crime scene. It tells you, "Hey, this gene is likely involved in this disease!" This clue is good. It doubles the chances of a drug succeeding compared to guessing.

But sometimes, the fingerprint is blurry. It points to a whole neighborhood of genes, and it's hard to know exactly which house the culprit lives in.

The New Super-Clue: Proteogenomics

This paper introduces a new, sharper clue called Proteogenomics (specifically looking at plasma proteins).

If Human Genetic Evidence is a fingerprint, think of Proteogenomics as a live video feed of the suspect's actions. It doesn't just tell you where the gene is; it shows you exactly how that gene changes the levels of specific proteins in your blood, and how those protein changes directly cause the disease.

The Big Discovery

The researchers asked: "What happens if we use both the fingerprint (genetics) and the video feed (proteomics) together?"

The Answer: It's a game-changer.

• Genetics alone: Makes a drug 2.6 times more likely to succeed.
• Genetics + Proteomics: Makes a drug 4.7 times more likely to succeed.

That's like upgrading from a bicycle to a rocket ship. The combination of these two clues makes the path to a successful medicine much clearer.

How They Did It (The Recipe)

The team didn't just guess; they cooked up a massive recipe using data from thousands of people:

1. The Ingredients: They gathered data on 8,000 different diseases and looked at how genes affect protein levels in the blood.
2. The Test: They used a statistical method called "Mendelian Randomization" (think of it as a natural experiment) to see if changing a protein level would actually change the disease risk.
3. The Match: They compared their predictions against a giant list of real-world drug trials (over 29,000 drug attempts) to see which ones actually made it to the finish line (getting approved by the FDA).

The "Missing Pieces" Puzzle

One of the coolest parts of the study is how it helps with the "blind spots."

Imagine you are looking for a specific type of fish in a lake.

• Old Method (Genetics only): You have a map that says the fish is in the lake, but the map is fuzzy. You might miss the fish if it's hiding in the weeds.
• New Method (Proteomics): You have a sonar that actually hears the fish.

The study found that for certain types of proteins (like enzymes and kinases), the old map was almost useless. But when they added the sonar (proteomics), they suddenly found a huge number of successful drug targets that were previously invisible.

However, there is a catch: The "sonar" they used (current blood tests) isn't perfect yet. It's great at hearing some fish, but it's bad at hearing others (like certain receptors and channels). The authors say, "We need better sonars to hear the fish we're currently missing!"

Real-World Examples

• The TNF Example: There was a famous drug for a joint disease (Ankylosing Spondylitis) that worked because it blocked a protein called TNF. But the old genetic maps missed this connection because the area of DNA involved was too messy to read. The new proteomic "video feed" saw it clearly.
• The CSF3 Example: Sometimes the "video feed" can be tricky. A signal suggested a drug would cause a disease, but when they looked closer at the specific DNA location, they realized the signal was actually a complex feedback loop, not the direct cause. This teaches us to be careful and look at the whole picture, not just one angle.

The Bottom Line

This paper tells us that combining genetic maps with protein data is the future of drug discovery.

It's like trying to navigate a city. If you only have a street map (genetics), you might get lost in traffic. But if you have the map plus a real-time GPS showing traffic jams and road closures (proteomics), you are much more likely to reach your destination (a successful, life-saving medicine) without crashing.

In short: Adding protein data to genetic data doesn't just help a little; it nearly doubles the success rate of new drugs, saving time, money, and potentially millions of lives.

Technical Summary

1. Problem Statement

Drug development suffers from high attrition rates, with suboptimal therapeutic efficacy being a primary cause of failure. While human genetic evidence (e.g., Genome-Wide Association Studies or GWAS) has been shown to double the probability of clinical success for drug targets, GWAS signals often span multiple genes, making it difficult to confidently assign causal genes.

• The Gap: Previous studies established that integrating cis-QTLs (including pQTLs and eQTLs) improves success rates, but it remained unclear whether plasma protein quantitative trait loci (pQTLs) specifically provide added value over standard human genetic evidence (like Locus-to-Gene scores) in prioritizing therapeutic hypotheses.
• The Objective: To quantify the impact of pQTL-supported evidence on the clinical success of drug target–indication (T–I) pairs, specifically determining if pQTLs complement existing genetic evidence to further increase the likelihood of a drug advancing from Phase I to launch.

2. Methodology

The authors conducted a large-scale retrospective enrichment analysis integrating proteogenomic data with clinical trial outcomes.

• Data Integration:
  • Proteogenomic Data: Aggregated eight publicly available plasma proteogenomic datasets (including Olink and SomaScan platforms) to generate genome-wide pQTLs.
  • GWAS Traits: Mapped to ~8,000 complex traits from the GWAS Catalog, Pan-UK Biobank, and FinnGen.
  • Clinical Outcomes: Utilized a curated dataset of 29,476 drug target–indication (T–I) pairs from Citeline Pharmaprojects (Minikel et al.), tracking their progression from Phase I to launch.
• Hypothesis Generation & Filtering:
  • Mendelian Randomization (MR): Performed 47.2 million MR tests to link genetically predicted plasma protein levels to GWAS traits.
  • Locus-to-Gene (L2G) Scoring: Used Open Targets Genetics (v8) L2G scores. A T–I pair was considered "supported" if it met:
    • MeSH Similarity: $\ge$ 0.8 (ensuring the disease indication matches the target).
    • L2G Share: $\ge$ 0.5 (indicating high confidence in the causal gene assignment).
    • pQTL Support: MR p-value passed a strict Bonferroni threshold ($p \le 1.06 \times 10^{-9}$) and, for a stricter subset, colocalization support ($H_4 \ge 0.8$).
• Statistical Analysis:
  • Calculated Relative Success (RS): The ratio of success rates (Phase I to Launch) for genetically supported pairs versus unsupported pairs.
  • Sensitivity Analyses: Tested for confounding by therapeutic area, analyzed different MR-coloc subtypes (cis, trans, mixed), and evaluated enrichment across protein families.
  • Tools: Used GSMR for MR, TOP-LD for variant annotation, and R (v4.3.3) for statistical modeling.

3. Key Contributions

• Quantification of Added Value: Demonstrated that pQTL evidence provides a distinct, additive layer of validation beyond standard genetic evidence (L2G scores).
• Protein Family Insights: Identified that pQTL support significantly boosts success rates for specific druggable families (e.g., enzymes, kinases) that show little enrichment from genetic evidence alone.
• Complementarity with L2G: Showed that pQTLs can rescue therapeutic hypotheses where L2G scores are low or absent (e.g., due to complex regions like HLA/MHC being excluded from standard pipelines).
• Directionality Warning: Highlighted the risk of relying solely on trans-pQTLs for therapeutic directionality, as they may reflect feedback loops rather than causal mechanisms, necessitating integration with cis-evidence.
• Resource Release: Provided a browsable database of pQTL-supported target-trait pairs and code repositories for reproducibility.

4. Key Results

• Dramatic Increase in Clinical Success:
  • pQTL + Genetic Evidence: T–I pairs supported by both pQTLs and human genetic evidence were 4.7 times more likely to advance from Phase I to launch (RS = 4.73; 95% CI: 3.51–6.36).
  • Genetic Evidence Alone: Human genetic evidence (L2G $\ge$ 0.75) without pQTL support showed a 2.6-fold increase (RS = 2.60).
  • Combined High Confidence: Pairs with L2G $\ge$ 0.75 plus pQTL support achieved an RS of 5.65, more than doubling the success rate of genetic evidence alone.
• Robustness Across Subtypes: Enrichment was consistent across MR-coloc subtypes (cis, trans, mixed), though most successful pairs were supported by multiple evidence streams (intersection of up to four types).
• Protein Family Specificity:
  • Enriched Families: Enzymes and kinases showed massive RS increases when pQTLs were added.
  • Gaps: Families like GPCRs, nuclear receptors, and ion channels showed null enrichment, attributed to inadequate assay coverage in current proteomic panels (Olink/SomaScan) rather than a lack of biological validity.
• Case Studies:
  • TNF/Ankylosing Spondylitis: pQTLs validated the association despite the exclusion of the HLA region in standard L2G pipelines.
  • SOST/Osteoporosis: MR evidence supported the target despite modest L2G scores.
  • CSF3R/Neutropenia: A trans-signal suggested a misleading therapeutic direction, which was corrected only when considering cis-acting variants, highlighting the need for context-aware interpretation.

5. Significance and Limitations

Significance:

• Strategic Prioritization: The study provides strong empirical evidence that integrating proteogenomic data (pQTLs) with standard genetic evidence is a superior strategy for de-risking drug discovery pipelines.
• Panel Expansion: It identifies a critical need to expand proteomic panels to cover under-represented protein families (GPCRs, transporters) to unlock further therapeutic opportunities.
• Validation Framework: It establishes a framework where pQTLs act as a "second opinion" to confirm causal genes, particularly in complex genomic regions where standard GWAS-to-gene mapping fails.

Limitations:

• Assay Coverage: The current success estimates are driven by only 12 targets across 23 launched pairs, heavily dependent on the coverage of existing proteomic platforms.
• Variant Interpretation: Some cis-pQTLs may be driven by protein-altering variants (PAVs) that affect antibody/aptamer binding, potentially complicating the interpretation of effect direction.
• Population Generalizability: Analyses rely primarily on European-ancestry cohorts and circulating plasma proteins, which may not fully reflect tissue-specific disease mechanisms or non-European populations.
• eQTL Exclusion: The study did not assess eQTL-based enrichment due to resource limitations and comparability issues with the genome-wide pQTL framework.

Conclusion:
The integration of plasma proteogenomics with human genetic evidence significantly enhances the prediction of clinical drug success. By moving beyond simple genetic association to include functional protein-level evidence, researchers can prioritize therapeutic hypotheses with higher confidence, particularly for targets where genetic signals are ambiguous or where standard pipelines miss causal genes.