Genetic similarity among 178 disease phenotypes predicts therapeutic and side effects for 1,711 drugs

This study demonstrates that a gene-agnostic method quantifying genetic similarity among 178 diseases can successfully predict novel therapeutic uses and adverse effects for 1,711 drugs, offering a new pathway for drug discovery that bypasses the need for identifying specific causal genes or targets.

The Gist

Imagine you have a massive library of human health information. Inside this library, every disease has a "genetic fingerprint" made up of thousands of tiny DNA clues. For a long time, scientists tried to find new medicines by looking for a single, perfect match between a drug's target and a specific disease's fingerprint. But this is like trying to find a needle in a haystack, and often, the needle is hidden or missing.

This paper proposes a different way to look at the library. Instead of hunting for a single perfect match, the authors ask: "Which diseases look genetically similar to each other, even if they seem totally different on the surface?"

Here is the breakdown of their work using simple analogies:

1. The Core Idea: The "Genetic Cousin" Theory

Think of diseases as people. Some diseases are obvious relatives (like a heart attack and high blood pressure—they both affect the heart). But this study looked for "genetic cousins"—diseases that might live in different neighborhoods (different body systems) but share the same family DNA.

The researchers hypothesized that if Disease A and Disease B are genetic cousins, a medicine that works for Disease A might also work for Disease B, even if no one realized they were related before.

2. The Method: Five Different "Rulers"

To measure how similar two diseases are, the team didn't just use one ruler; they built five different measuring tools.

• Tool 1 & 2: Look at the big picture of the whole genome (like comparing the overall height and build of two people).
• Tool 3 & 4: Look at specific genes and how they are turned on or off in different tissues (like comparing their specific hobbies or skills).
• Tool 5: Looks at how genes and diseases overlap in specific molecular neighborhoods.

They used these tools to compare 178 different diseases.

3. The Discovery: Similarity Predicts Success

They tested their theory against a list of 1,711 common drugs. They asked: "Do diseases that are genetically similar actually share the same drugs?"

The answer was a resounding yes.

• The "Indication" Test: If a drug treats Disease A, and Disease B is genetically similar to Disease A, the drug is more likely to treat Disease B too.
• The "Side Effect" Test: If a drug causes a side effect in Disease A, and Disease B is genetically similar, the drug is more likely to cause that same side effect in Disease B.

4. The "Magic" Prediction Model

The authors built a computer model (a "predictor") that combines all five measuring tools. They used this model to guess new uses for drugs.

• For New Treatments: When the model gave a prediction a high score (probability > 0.1), those drug-disease pairs were twice as likely to succeed in clinical trials and get approved by regulators compared to random guesses.
• For Side Effects: When the model predicted a side effect with a high score (probability > 0.2), it was 1.4 times more likely to be a real side effect.

5. The "Two-Sided Coin" Surprise

One of the most interesting findings is that the model works both ways.

• The model trained to find new cures was surprisingly good at predicting side effects.
• The model trained to find side effects was surprisingly good at predicting new cures.

The Analogy: Imagine a drug is a key. The "lock" it opens is the disease it cures. But sometimes, that same key accidentally jams a different lock in the house, causing a side effect. The study found that the genetic blueprint of the "cured" lock and the "jammed" lock are often so similar that if you know one, you can guess the other.

6. Real-World Examples from the Paper

The paper gives two concrete examples of how this works:

• Naltrexone (Alcoholism drug) $\rightarrow$ IBS: Naltrexone treats alcoholism. The model predicted it might help Irritable Bowel Syndrome (IBS). Why? Because IBS is genetically similar to alcoholism in this specific context. The biology of the gut and the brain's opioid receptors are linked, even though they seem like different body parts.
• Anastrozole (Breast cancer drug) $\rightarrow$ Malabsorption: Anastrozole lowers estrogen. The model predicted it might cause "malabsorption syndrome" (trouble absorbing nutrients). Why? Because the drug's known side effects (bone loss, vein inflammation) share a genetic link with how the gut absorbs nutrients.

What This Means (According to the Paper)

The authors emphasize that this method is gene-agnostic. This means you don't need to know exactly which gene causes a disease to find a cure. You just need to know that Disease A and Disease B share a genetic "vibe."

Important Limitations Mentioned:

• The paper admits this doesn't work for every drug. Some drugs just treat symptoms (like a painkiller for a headache) rather than fixing the underlying genetic cause, so this method might miss those.
• The data isn't perfect; some diseases have more genetic data than others, which can make the "rulers" less accurate for certain conditions.

In Summary:
This paper shows that by looking at the genetic "family trees" of diseases, we can predict which drugs will work for new diseases and which will cause new side effects. It's like realizing that two houses look different on the outside, but because they were built with the same blueprints, the same repairman (drug) can fix both of them.

Technical Summary

1. Problem Statement

While human genetics (via GWAS) has shown promise in identifying drug targets, its clinical translation is hindered by the difficulty of linking genetic variants to specific causal genes. Traditional approaches often discard a significant fraction of genetic data because they rely on identifying genome-wide significant loci and their causal genes. Furthermore, existing methods often fail to leverage the phenomenon of pleiotropy (where genetic variants influence multiple traits) to its full potential for drug repurposing, particularly for diseases that appear phenotypically unrelated. There is a need for a gene-agnostic method that utilizes the broad spectrum of genetic data to predict novel drug-disease links (both therapeutic indications and adverse side effects) without requiring the prior identification of specific causal genes.

2. Methodology

The authors developed a framework to quantify genetic similarity between disease pairs and use this metric to predict drug-disease associations.

A. Data Collection & Curation

• Diseases: 178 distinct disease phenotypes.
• Drugs: 1,711 common drugs.
• Phenotypic Control: To isolate genetic signals from clinical similarities, the authors used disease annotations (ICD10, UK HRCS) and ClinGraph (a clinical knowledge graph) to calculate phenotypic similarity embeddings. Pairs with high phenotypic similarity were excluded in specific analyses to ensure the model relied on genetic, not just clinical, overlap.

B. Five Genetic Similarity Metrics
The study constructed five distinct metrics to capture different layers of genetic architecture:

1. LDSC-based: Genome-wide genetic correlation using LD Score Regression (from GWAS Atlas).
2. MAGMA-based: Gene-level association p-values aggregated from GWAS signals (GWAS Atlas).
3. S-MultiXcan-based: Gene regulation profiles based on genetically predicted gene expression across 44 tissues (PhenomeXcan).
4. L2G-based: Machine learning-derived probabilities of causal genes at GWAS loci (Open Targets).
5. COLOC-based: Colocalization of GWAS loci with molecular QTLs (eQTLs, pQTLs, sQTLs) (Open Targets).

C. Predictive Modeling

• Approach: A Bayesian logistic regression model was developed to integrate all five genetic similarity metrics. This allows for predictions even when data is missing for specific metric-disease pairs.
• Input: For a given drug-disease pair, the model calculates the maximum genetic similarity between the candidate disease and the drug's known indications (or known side effects).
• Training: Models were trained on known drug-disease pairs. Crucially, the authors trained separate models: one using all pairs and another restricted to phenotypically dissimilar pairs to prove the utility of genetic similarity independent of clinical co-occurrence.
• Validation: Predictions were validated against:
  • Clinical trial progression data (Phase I to Approval).
  • Independent side effect databases (SIDER, offSIDES).
  • Permutation tests to assess significance against random chance.

3. Key Contributions

• Gene-Agnostic Framework: The study proposes a method that bypasses the need to identify specific causal genes, leveraging aggregate genetic similarity instead.
• Phenome-Wide Scale: It evaluates 178 diseases and 1,711 drugs, providing a comprehensive map of potential drug-disease links.
• Decoupling Genetics from Phenotype: By rigorously stratifying for phenotypic similarity, the authors demonstrate that genetic similarity provides predictive power even for diseases that do not share obvious clinical features.
• Bidirectional Prediction: The study demonstrates that genetic similarity can predict both therapeutic indications (repurposing) and adverse side effects, and notably, that these two domains share a genetic basis.

4. Key Results

• Genetic Similarity Correlates with Drug Sharing: Diseases with higher genetic similarity scores share significantly more drugs (both as indications and side effects), even after controlling for phenotypic similarity and body system.
• Predictive Power for Indications:
  • The model achieved an AUROC of 0.747 on held-out data (all pairs) and 0.559 when restricted to phenotypically dissimilar pairs.
  • Clinical Trial Success: Drug-disease pairs with a predicted probability >0.1 were 2.03 times more likely to progress from Phase I clinical trials to regulatory approval compared to pairs with low predicted probability.
• Predictive Power for Side Effects:
  • The side effects model achieved an AUROC of 0.548 on phenotypically dissimilar pairs.
  • Predictions with probability >0.2 were 1.42 times more likely to correspond to true adverse effects (validated against SIDER).
• Shared Genetic Basis (Cross-Prediction):
  • The Indication Model could predict true side effects better than chance (AUROC = 0.525, p < 0.001).
  • The Side Effect Model could predict true indications better than chance (AUROC = 0.542, p < 0.001).
  • This implies that the biological pathways driving a drug's efficacy often overlap with those causing its toxicity.
• Case Studies:
  • Naltrexone for IBS: Predicted (prob=0.283) based on genetic similarity to alcoholism. Supported by biological rationale (opioid receptors in the GI tract) despite lacking direct GWAS hits for IBS targets.
  • Anastrozole for Malabsorption: Predicted (prob=0.61) based on similarity to osteoporosis/phlebitis, linked via estrogen signaling pathways affecting the gut.

5. Significance and Implications

• Accelerated Drug Discovery: This method identifies repurposing candidates that would be missed by approaches relying solely on genome-wide significant GWAS hits or causal gene identification. It is particularly valuable for diseases with weak or missing GWAS signals at specific drug targets.
• Safety Profiling: The ability to predict side effects from genetic similarity to known adverse events offers a proactive tool for safety evaluation, potentially reducing clinical trial attrition.
• Mechanistic Insight: The finding that indications and side effects share a genetic basis suggests that drug efficacy and toxicity are two sides of the same biological coin, driven by shared pleiotropic pathways.
• Resource for Validation: The authors provide a public repository of predicted probabilities for thousands of drug-disease pairs, serving as a prioritization resource for experimental validation.

Limitations Noted: The study acknowledges that symptomatic treatments may not be captured by genetic associations, data incompleteness across different genetic databases, and the varying sample sizes of GWAS studies affecting confidence in similarity estimates. However, the rigorous control for phenotypic similarity strengthens the validity of the genetic signal identified.