A DNA foundation model predicts osteoporosis risk genes without proximity bias

This paper introduces Rosalind, a DNA foundation model that overcomes the proximity bias of traditional gene-mapping approaches by accurately predicting distal variant-gene regulatory relationships from sequence, thereby successfully identifying novel osteoporosis risk genes and demonstrating a scalable framework for translating genetic insights into drug discovery.

The Gist

The Big Problem: The "Nearest Neighbor" Mistake

Imagine you are looking at a massive, crowded city (the human genome). You spot a strange graffiti tag (a genetic mutation) on a wall. You want to know: Which building is this graffiti actually affecting?

For decades, scientists have used a very simple rule to answer this: "It must be the building right next to the wall."

This is called the "Nearest Neighbor" rule. If a mutation is found near Gene A, scientists assumed Gene A was the culprit causing the disease.

The Problem: DNA isn't a straight line; it's a tangled ball of yarn inside a tiny cell. A mutation might be physically close to Gene A, but because of how the DNA folds, it might actually be touching and controlling Gene Z, which is miles away in the linear sequence. Relying only on distance is like assuming the person shouting at you is the one standing right next to you, ignoring the fact that they might be shouting through a megaphone to someone across the street.

This "proximity bias" has caused scientists to miss the real causes of diseases like osteoporosis (brittle bones).

The Solution: Meet "Rosalind"

The authors of this paper built a new AI tool called Rosalind. Think of Rosalind not as a rule-follower, but as a super-smart detective who understands the "language" of DNA.

• How it works: Instead of just measuring distance, Rosalind reads the DNA sequence like a sentence. It understands the grammar, the punctuation, and the long-range connections. It knows that a word (mutation) at the beginning of a paragraph can change the meaning of a word at the very end, even if they are far apart.
• Training: Rosalind was trained on a massive library of human genetic data (from the GTEx project) to learn which mutations actually change how genes behave.

The Test: The "Bone Builder" Experiment

To prove Rosalind was better than the old "Nearest Neighbor" rule, the team used Osteoporosis (brittle bones) as a test case.

1. The Setup: They took 1,103 known genetic "clues" linked to weak bones.
2. The Prediction:
   • The Old Way (Nearest Neighbor) pointed to the genes closest to the clues.
   • Rosalind pointed to different genes, often ones far away from the clue.
3. The Lab Test: They took human bone-building cells (osteoblasts) and used a molecular pair of scissors (CRISPR) to "turn off" the genes Rosalind predicted.
   • The Result: When they turned off the genes Rosalind picked (the "distal" ones), the bone cells stopped building bone properly.
   • The Surprise: When they turned off the "nearest neighbor" genes, the bone cells kept working fine!

The Analogy: Imagine you are trying to fix a broken car engine. The old rule says, "Check the part closest to the broken wire." Rosalind says, "No, check the fuel pump three feet away." The team went and checked the fuel pump, and that was the problem.

Why This Matters

1. Better Drug Discovery: If you are trying to invent a drug to cure a disease, you need to target the right gene. If you target the wrong one (because you followed the "nearest neighbor" rule), the drug will fail. Rosalind helps find the real target, saving time and money.
2. New Insights: Rosalind found that genes involved in primary cilia (tiny antenna-like structures on cells that sense movement) are crucial for bone health. This is a new discovery that the old methods completely missed.
3. Scalability: This isn't just for bones. Because Rosalind understands the "language" of DNA, it can be applied to diabetes, heart disease, asthma, and more to find the true causes.

The Bottom Line

This paper introduces a new AI tool that stops guessing based on distance and starts understanding the complex, 3D reality of our DNA. By doing so, it helps scientists find the true "villains" behind diseases, leading to better medicines and cures.

In short: We stopped looking at the person standing next to the crime scene and started looking at who actually pulled the trigger, even if they were hiding in the next room.

Technical Summary

1. Problem Statement

• High Attrition in Drug Development: Approximately 90% of drug candidates fail. Human genetic evidence is a robust predictor of success, yet integrating it into target identification remains challenging.
• The Variant-to-Gene (V2G) Bottleneck: While Genome-Wide Association Studies (GWAS) identify disease-associated variants, ~90% of these lie in non-coding regions. Linking these regulatory variants to their specific effector genes (eGenes) is difficult.
• Proximity Bias: Current V2G mapping methods (e.g., Open Targets L2G) rely heavily on the heuristic that the "nearest gene" to a variant is the causal one. This ignores the 3D architecture of the nucleus, where enhancers can regulate genes hundreds of kilobases away. This bias leads to missed causal targets and false positives in complex genomic loci.
• Limitations of Existing Models: While deep learning models like Enformer can capture long-range interactions, they are typically trained on reference genomes and struggle to generalize to individual-level genetic variation or predict genotype-specific expression changes without fine-tuning.

2. Methodology

The authors introduce Rosalind, a DNA foundation model designed to predict causal variant-gene relationships directly from sequence without proximity assumptions.

• Architecture:
  • Based on the Enformer architecture (a Transformer-based model with a large receptive field).
  • Key Modification: Replaced relative positional encodings with Rotary Positional Embeddings (RoPE) to better encode long-range distances.
  • Fine-tuning Strategy: The model was fine-tuned on a tissue-agnostic framework using ~17,000 variant-gene pairs derived from fine-mapped GTEx cis-eQTL data.
  • Training Task: A binary classification task where the model learns to distinguish causal variants (Posterior Inclusion Probability, PIP > 0.9) from non-causal ones (PIP < 0.01). It processes both reference and alternate alleles to generate allele-specific embeddings, which are fed into a lightweight Multi-Layer Perceptron (MLP) discriminator.
• Benchmarking:
  • Local Scale: Tested against GeneGenie (a local promoter model) on yeast promoter screens and human osteoblast enhancer assays.
  • Global Scale: Compared against Enformer (adapted for CAGE signal prediction) and distance-based baselines on held-out GTEx data.
  • GWAS Application: Applied to four complex traits (Type 2 Diabetes, Hypertension, Asthma, Psoriasis) and Osteoporosis, comparing predictions against Open Targets L2G scores.
• Experimental Validation (Osteoporosis Case Study):
  • Target Selection: Applied Rosalind to 1,103 lead variants from an estimated Bone Mineral Density (eBMD) GWAS.
  • Assay: An arrayed CRISPR/Cas9 knockout screen in human hFOB1.19 osteoblasts.
  • Readout: Quantification of hydroxyapatite mineralization (a proxy for bone formation) using Alizarin Red staining.
  • Design: For each variant, the model-predicted "distal" gene and the "nearest" gene were knocked out side-by-side to determine which influenced the phenotype.

3. Key Contributions

• Rosalind Model: A novel DNA foundation model that successfully fine-tunes on genetic variation to predict regulatory impacts, overcoming the limitations of reference-genome-only models.
• Overcoming Proximity Bias: Demonstrated that deep learning models can accurately identify distal effector genes, a capability lacking in current standard-of-care tools like L2G.
• Translational Validation: Provided the first large-scale experimental validation showing that distal gene predictions from a deep learning model are functionally more relevant to disease phenotypes than nearest-gene predictions.
• Novel Biological Insight: Identified a previously underappreciated role for primary cilia structure genes (e.g., CATIP, GANAB) in osteoporosis risk, which were missed by proximity-based methods.

4. Key Results

• Benchmarking Performance:
  • Rosalind outperformed Enformer and distance-based baselines in identifying causal variant-gene pairs, particularly for distal genes.
  • In the GTEx held-out set, Rosalind maintained higher AUROC/AUPRC scores as genomic distance increased, whereas performance for nearest-gene baselines dropped significantly.
• Generalizability Across Traits:
  • Across four complex traits, Rosalind prioritized distal genes at a rate >2x higher than Open Targets L2G (which is dominated by nearest-gene assignments).
  • Rosalind successfully recovered known drug targets missed by L2G (e.g., GIPR and NDUFA7 for diabetes, GUCY1B1 for hypertension) that were distal to the lead variants.
• Osteoporosis Experimental Validation:
  • Rosalind identified 239 high-confidence disease risk genes (DRGs), the majority of which were distal.
  • Functional Hit Rate: In the CRISPR screen, Rosalind-predicted distal genes were significantly more likely to alter osteoblast mineralization than the nearest genes (McNemar's test, p = 0.011).
  • Specific Findings:
    • Validated known pathways (e.g., TNFRSF11A/B, COL1A1/2, WNT1).
    • Identified FDPS (target of bisphosphonates) as the effector for a variant where the nearest gene was THBS3.
    • Identified CHUK (NF-κB regulator) as the effector for a variant where the nearest gene was PKD2L1.
    • Discovered that genes involved in primary cilia structure (CATIP, GANAB) were functional hits in the assay, suggesting a novel mechanistic link between cilia and bone density.

5. Significance

• Paradigm Shift in Target Discovery: The study establishes that sequence-based foundation models can replace heuristic proximity rules, offering a scalable, generalizable framework for translating non-coding GWAS signals into actionable drug targets.
• Drug Development Impact: By identifying causal genes that are currently ignored by proximity-based methods, Rosalind can uncover novel therapeutic targets, potentially increasing the success rate of clinical trials.
• Biological Insight: The validation of distal genes in osteoblasts confirms that regulatory logic in complex disease is often non-local, and that deep learning can decode these long-range interactions without requiring expensive, tissue-specific functional genomics datasets (like sc-eQTLs) for every new disease.
• Scalability: The approach provides a blueprint for applying DNA language models to any complex trait, accelerating the "genetics-to-drug" pipeline.