Identifying severe COVID-19 risk variants modulating enhancer reporter activity in lung cells

Using a massively-parallel reporter assay and deep learning models in lung epithelial cells, this study identifies 29 severe COVID-19 risk variants that allele-specifically modulate enhancer activity, revealing their potential to alter gene expression and influence disease outcomes.

The Gist

The Big Picture: Why Do Some People Get Sicker?

Imagine the SARS-CoV-2 virus (which causes COVID-19) is a burglar trying to break into a house (your body). Most houses have standard locks, but some houses have slightly different locks or weak spots in the walls.

Scientists already knew that some people are genetically more likely to get severely sick from this "burglar." They found thousands of tiny differences in people's DNA (the instruction manual for building a human) that seem to be linked to severe illness. These differences are like typos in the instruction manual.

The Problem: Most of these typos aren't in the parts of the manual that build the actual furniture (proteins). They are in the "marginal notes" or "sticky notes" that tell the factory when and how much to build. We knew the typos existed, but we didn't know which specific typo was causing the problem, or what it was actually breaking.

The Experiment: The "DNA Test Kitchen"

To solve this, the researchers built a massive "test kitchen" to see which typos actually change how the factory works.

1. The Ingredients (The Library): They took 4,894 specific DNA "typos" found in people who got very sick with COVID. They also looked at combinations of typos that often travel together (like a set of matching keys).
2. The Test (STARR-seq): They put these DNA snippets into a special cell line (A549) that acts like a model of the lungs. Think of these cells as little factories.
   • They attached each DNA snippet to a lightbulb.
   • If the DNA snippet is a "bad" instruction that messes up the factory, the lightbulb might flicker, get brighter, or go dark.
   • If it's just a harmless typo, the lightbulb stays normal.
3. The Scale: They didn't test them one by one. They tested all 4,894 at once, like throwing thousands of different keys into a giant lockbox to see which ones turn the tumblers.

The Results: Finding the Culprits

Out of the thousands of typos they tested, they found 29 specific "bad actors" that actually changed how the lung cells behaved.

• The "Volume Knobs": Imagine a volume knob on a radio. Some of these typos turned the volume down (making the cell's defense system quieter), while others turned it up (making it too loud).
• The "Double Trouble": They also tested pairs of typos. Sometimes, two typos together were worse than just adding their effects up. It's like two people whispering in a room; individually, you can't hear them, but together, they create a distraction that stops the whole conversation.
• The Location: Many of these bad typos were found in the Lungs, specifically in the parts of the DNA that control how the body fights viruses (like the Interferon system, which is the body's "alarm system").

The Detective Work: AI as a Sidekick

The researchers also used advanced Artificial Intelligence (Deep Learning models) to try and predict which typos were bad before they did the experiment.

• The Reality Check: The AI was good at guessing the big, obvious problems, but it missed many of the subtle ones. It's like a weather forecast that predicts a hurricane but misses a sudden, dangerous thunderstorm.
• The Lesson: You can't just rely on the computer guess; you still need to do the real experiment (the "test kitchen") to be sure. However, once they found the bad typos, the AI helped explain why they were bad (e.g., "This typo broke the binding site for a specific protein").

Why Does This Matter? (The "So What?")

This study is like finding the specific broken gears in a clock rather than just saying "the clock is broken."

1. New Drug Targets: Now that we know which genes are being messed up (like IFNAR2, a key part of the immune alarm, or CRHR1, which relates to stress and lung repair), drug companies can design medicines to fix those specific gears.
2. Understanding Severity: It explains why some people have mild cases and others need intensive care. It's not just bad luck; it's specific genetic "typos" in their lung cells.
3. Future Research: This gives scientists a "hit list" of the most important DNA variations to study further. They can now use these findings to develop better treatments or even personalized medicine for people with these specific genetic risks.

In a Nutshell

Think of the human genome as a massive, complex instruction manual for building a body. This paper took a giant list of "suspected typos" found in people who got very sick with COVID, tested them in a lung cell lab, and identified the 29 specific typos that actually break the instructions. They also showed that sometimes, two typos working together cause more damage than the sum of their parts. This helps us understand the "why" behind severe COVID-19 and points the way toward better cures.

Technical Summary

1. Problem Statement

While Genome-Wide Association Studies (GWAS) have identified thousands of genetic variants associated with severe COVID-19 outcomes, the vast majority reside in the non-coding genome. The specific causal variants, their functional mechanisms, and the target genes they regulate remain largely unknown due to Linkage Disequilibrium (LD), where non-causal and causal variants are co-inherited. Furthermore, regulatory elements (enhancers) often function in a tissue-specific manner, and previous functional studies have been limited to small numbers of loci or non-relevant cell types (e.g., erythroleukemia cells). There is a critical need for high-throughput, lung-specific functional validation of these variants to elucidate disease mechanisms and identify therapeutic targets.

2. Methodology

The authors employed a massively parallel reporter assay (MPRA) using STARR-seq (Self-Transcribing Active Regulatory Region sequencing) to screen thousands of variants in a lung-relevant context.

• Variant Library Design:

  • Source: Variants were collated from the GenOMICC GWAS studies (2nd and 3rd releases), focusing on severe cases requiring intensive care.
  • Selection: Included 2,528 fine-mapped variants (99% credible sets), 1,465 variants in high LD ($r^2 > 0.7$) with lead variants, and 901 rare variants (MAF > 0.02%).
  • Total Variants: 4,894 unique variants (mostly SNPs).
  • Combinatorial Design: To test for non-additive effects, the library included all possible allelic combinations for 777 variants located within 100 bp of each other, generating 3,776 combinatorial oligonucleotides.
  • Controls: Included positive controls (known active enhancers in A549) and negative controls (scrambled sequences).
  • Oligonucleotides: 170-bp sequences centered on each variant (reference and alternate alleles) flanked by adapters.

• Experimental Workflow:

  • Cell Line: A549 lung adenocarcinoma cells (a model for type II alveolar epithelial cells).
  • Assay: The oligonucleotide library was cloned into an hSTARR vector, transfected into A549 cells, and subjected to STARR-seq.
  • Sequencing: Input (DNA) and Output (RNA) libraries were sequenced in five biological replicates.
  • Analysis: Activity was measured as the log2 fold-change (log2FC) of RNA/DNA read counts. Variants were considered "active" if log2FC > 1 (FDR < 0.01). Allele-specific effects were identified using the mpralm linear model.

• Computational Integration:

  • Deep Learning: Two models were used to interpret results:
    • AlphaGenome: A multi-modal model predicting effects on chromatin accessibility, histone marks, and TF binding.
    • Malinois: A task-specific CNN trained on A549 MPRA data.
  • In-silico Mutagenesis (ISM): Used to identify specific transcription factor (TF) motifs disrupted or created by variants.
  • Data Integration: Overlapped results with ENCODE A549 datasets (ATAC-seq, DNase-seq, ChIP-seq) and GTEx eQTL/sQTL data.

3. Key Contributions

• Scale: The first large-scale functional screen of >4,800 severe COVID-19 risk variants specifically in lung epithelial cells.
• Combinatorial Analysis: Systematic testing of variant pairs to distinguish between additive and non-additive (interdependent) regulatory effects.
• Deep Learning Validation: A critical evaluation of state-of-the-art deep learning models (AlphaGenome, Malinois) against experimental MPRA data, highlighting their strengths in hypothesis generation but limitations in detecting weak, allele-specific effects.
• Prioritization: Identification of a high-confidence set of causal variants and candidate target genes with direct relevance to lung pathology and immune response.

4. Key Results

• Identification of Active Variants:

  • Out of 4,894 single variants tested, 166 resided in sequences with enhancer activity (log2FC > 1).
  • Of these, 29 variants showed significant allele-specific activity (amVars), meaning the risk allele altered regulatory activity compared to the reference.
  • 22 variants decreased activity, and 7 increased activity.
  • Many amVars overlapped with endogenous active chromatin features (ATAC-seq, H3K27ac) in A549 cells.

• Combinatorial Effects:

  • Testing 3,776 combinatorial sequences revealed 16 variant pairs with significant activity.
  • Additivity: 56% (9/16) of pairs acted additively, where the combined effect matched the sum of individual effects.
  • Non-Additivity: Some pairs showed non-additive effects (e.g., loss of activity only when both variants were present), suggesting interdependent TF binding or synergistic repression.

• Deep Learning Performance:

  • AlphaGenome: Correctly predicted the direction of effect (gain/loss) for 31% of amVars but missed many true positives (high false-negative rate). It performed best for chromatin accessibility and histone marks.
  • Malinois: Showed reasonable overall performance (AUC = 0.73) but also struggled with weak enhancers.
  • Conclusion: While deep learning models cannot yet replace experimental screens for small-effect variants, they are valuable for generating hypotheses about mechanism (e.g., specific TF motifs).

• Biological Insights & Candidate Genes:

  • Interferon Signaling:
    • rs6517156 (IFNAR2): The most significant loss-of-activity variant. It disrupts a p53 motif and reduces IFNAR2 expression (an eQTL), impairing type I interferon response.
    • IFNA Cluster: Three rare variants in the IFNA gene cluster (chromosome 9) showed reduced activity, potentially affecting type I interferon production.
  • Viral Entry & Processing:
    • rs2297480 (FDPS): A gain-of-activity variant in the FDPS promoter region. It creates a G-rich motif and is an sQTL, potentially altering protein prenylation (mevalonate pathway) crucial for SARS-CoV-2 endolysosomal entry.
    • rs6471885 (RAB2A): An eQTL for increased RAB2A expression, a known risk factor for severe COVID-19.
  • Lung Damage & Repair:
    • CRHR1/KANSL1/MAPT Locus (Chr 17): Five prioritized variant pairs were found near CRHR1 (corticosteroid receptor), KANSL1, and MAPT. These are linked to lung fibrosis and corticosteroid response.
    • BMP2: A variant upstream of BMP2 (involved in pulmonary fibrosis) showed reduced activity.

5. Significance

This study provides a robust, experimentally validated resource for understanding the genetic architecture of severe COVID-19. By moving beyond statistical association to functional validation in a disease-relevant cell type, the authors:

1. Pinpointed Causal Variants: Identified specific non-coding variants that directly modulate enhancer activity in lung cells.
2. Elucidated Mechanisms: Demonstrated that disease risk can arise from both single variants and complex combinatorial interactions (additive and non-additive).
3. Highlighted Therapeutic Targets: Linked variants to pathways such as Type I interferon signaling, viral entry mechanisms, and lung fibrosis, suggesting potential targets for drug repurposing (e.g., corticosteroids, mevalonate pathway inhibitors).
4. Defined Limitations of AI: Provided empirical evidence that current deep learning models, while powerful, still require experimental validation to capture subtle, allele-specific regulatory effects in complex genomic contexts.

The identified variants serve as a prioritized list for future endogenous validation (e.g., prime editing) to confirm target genes and therapeutic potential.