The distribution of fitness effects of new mutations in regulatory regions of the D. melanogaster genome

By combining experimentally validated regulatory regions with realistic forward-in-time simulations, this study reveals that while population genetics methods struggle to infer the effects of mildly deleterious mutations, a large fraction of new mutations in *D. melanogaster* non-coding regions are moderately deleterious, ultimately demonstrating that these regions contribute the majority of new deleterious mutations and beneficial substitutions and are essential for accurately modeling background selection across the genome.

The Gist

The Big Picture: The "Silent" Parts of the Instruction Manual

Imagine the genome of a fruit fly (Drosophila melanogaster) as a massive, ancient instruction manual for building a living creature.

For a long time, scientists have been obsessed with the bold, capitalized words in this manual (the protein-coding genes). They know that if you change a bold word, the instructions might break completely, and the fly won't survive. These are the "strongly harmful" mutations.

However, this paper focuses on the small print, the footnotes, and the margins (the non-coding regulatory regions). These parts don't build the parts of the fly directly; instead, they act like the dimmer switches, volume knobs, and timers that tell the bold words when to turn on, how loud to be, and how long to stay on.

The big question the authors asked was: What happens if you scribble over these dimmer switches? Are the errors there just as deadly as errors in the bold words, or are they just annoying glitches?

The Investigation: Testing the "Dimmer Switches"

The researchers took a massive dataset of fruit flies from Africa (the ancestral home of the species, where the population is largest and most diverse). They wanted to map out the "Distribution of Fitness Effects" (DFE).

Think of the DFE as a weather map for mutations. It tells us:

• How many mutations are like a hurricane (strongly harmful)?
• How many are like a heavy rainstorm (moderately harmful)?
• How many are just a light drizzle (mildly harmful)?
• And how many are actually sunny days (beneficial)?

The Challenge: Finding the "Control Group"

To measure the damage of a scribble on a dimmer switch, you need to know what the "normal" weather looks like. In genetics, you need "neutral" sites—parts of the manual that don't do anything important, so you can see the background noise.

The problem? In the non-coding regions, it's hard to find a "blank page" that is truly blank. The authors had to be very careful to find spots that were truly neutral (like the space between paragraphs) to compare against the important regulatory spots.

The Findings: It's Not Just "Hurricanes"

Here is what they discovered, using our manual analogy:

1. Coding Regions (The Bold Words):
When you mess up a bold word, it's usually a catastrophe. About 65% of new mutations here are "hurricanes" (strongly deleterious). The fly usually can't function.

2. Regulatory Regions (The Dimmer Switches):
When you mess up a dimmer switch, it's rarely a total disaster. Instead, it's usually a moderate annoyance.

• The mutations here are mostly "moderately harmful" (like a rainstorm). They don't kill the fly immediately, but they make the fly slightly less efficient, slower, or less fertile.
• Because there are so many more dimmer switches than bold words in the genome, these "moderate rainstorms" actually make up the majority of all the bad mutations in the fly's genome.

3. The "Good" News (Beneficial Mutations):
Sometimes, changing a dimmer switch is an upgrade! The study found that while beneficial mutations are slightly less common in regulatory regions than in coding regions, they still happen frequently. In fact, because the regulatory regions are so vast, they contribute a huge chunk of all the "good" evolutionary changes the fly population makes.

Why This Matters: The "Background Noise" Problem

The authors also looked at Background Selection. Imagine you are trying to listen to a specific radio station (a neutral gene), but there is a lot of static (bad mutations) nearby. If the static is loud, it drowns out your signal.

• Old View: Scientists thought the static only came from the "bold words" (coding regions).
• New View: This paper shows that the "dimmer switches" (regulatory regions) are actually creating a huge amount of static.

If you ignore the regulatory regions, your map of the genome's "noise" is wrong. You can't accurately predict how much genetic diversity a population has unless you account for the fact that the "dimmer switches" are constantly being tweaked by natural selection.

The Takeaway

This paper is like realizing that while the engine block (coding DNA) is the most critical part of a car, the wiring harness and sensors (regulatory DNA) are actually where most of the wear and tear happens.

• Coding mutations are like a broken engine: rare, but fatal.
• Regulatory mutations are like frayed wires and loose sensors: very common, usually not fatal on their own, but they add up to make the car run less efficiently.

By understanding that these "fuzzy" regulatory regions are full of moderate problems, scientists can now build better models of how evolution works, how species adapt, and why genetic diversity looks the way it does. It turns out, the "small print" in the instruction manual is just as important as the big bold text.

Technical Summary

1. Problem Statement

The Distribution of Fitness Effects (DFE) of new mutations is a fundamental parameter in population genetics, governing the maintenance of genetic variation and shaping genome-wide patterns of polymorphism and divergence. While the DFE of protein-coding (nonsynonymous) mutations is well-characterized, the DFE of mutations in non-coding regulatory regions remains poorly understood.

Key challenges in inferring the DFE for non-coding regions include:

• Identification of Sites: Difficulty in distinguishing functionally important regulatory sites (selected) from putatively neutral sites, especially given the short length of regulatory elements.
• Genomic Architecture: The lack of "interdigitation" (interleaving) between selected non-coding sites and neutral sites, which is required for accurate inference when mutation and recombination rates are heterogeneous.
• Selective Interference: Potential bias introduced by Hill-Robertson interference (HRI) from linked selected sites in nearby coding regions.
• Data Scarcity: Regulatory regions constitute a small fraction of the genome, often providing insufficient data for robust statistical inference.

2. Methodology

The authors employed a multi-step approach combining empirical data analysis, forward-in-time simulations, and comparative genomics.

A. Data Curation and Annotation

• Populations: Utilized SNP data from three sub-Saharan African D. melanogaster populations (South/Zimbabwe, East/Rwanda-Uganda, West/Cameroon-Nigeria) from Coughlan et al. (2022). These populations were selected for minimal substructure and admixture.
• Selected Sites:
  • High-Confidence: Experimentally validated regulatory regions (enhancers, promoters, TFBS, miRNAs) from the REDfly database (0.2% of the genome).
  • Low-Confidence/Conserved: PhastCons elements (evolutionarily conserved across 27 insect species) and computationally predicted CRMs (62% of the genome).
  • Coding/UTR: Standard annotations from FlyBase.
• Neutral Sites: Identified putatively neutral intergenic sites by excluding all functional annotations. Further filtered using single-site phastCons scores (<0.1) to remove weakly selected sites. Crucially, neutral sites were restricted to 5kb flanking regions of selected sites to ensure similar mutation/recombination rate environments (interdigitation).

B. Simulation Framework

To validate inference accuracy, the authors used SLiM 4.0.1 for forward-in-time simulations mimicking D. melanogaster demography and genomic architecture:

• Realism: Incorporated empirically derived recombination maps (Comeron et al.), mutation rate heterogeneity, and exact genomic positions of selected elements.
• Scenarios: Simulated DFEs with varying proportions of weakly, moderately, and strongly deleterious mutations, as well as beneficial mutations (0.02%).
• Testing: Evaluated the performance of three inference tools (DFE-alpha, GRAPES, fastDFE) under varying numbers of sites and individuals to determine power and bias.

C. Empirical Inference

• Applied DFE-alpha, GRAPES, and fastDFE to the empirical data using folded and unfolded Site Frequency Spectra (SFS).
• Estimated the deleterious DFE (modeled as a Gamma distribution with mean $2N_es_d$ and shape $\beta$) and the proportion of beneficial substitutions ($\alpha$).
• Background Selection (BGS): Used Bvalcalc to generate genome-wide maps of expected neutral diversity ($B$) under purifying selection, comparing models using only coding regions vs. models including non-coding DFEs.

3. Key Contributions

1. Validation of Inference Methods: Provided a rigorous benchmark showing that while DFE inference is accurate for moderately and strongly deleterious mutations, weakly deleterious DFEs are prone to misinference (underestimation of weakly deleterious, overestimation of moderately deleterious) due to the difficulty in distinguishing them from neutral sites.
2. Characterization of Non-Coding DFE: Established that regulatory non-coding mutations are generally moderately deleterious (unlike coding mutations, which are predominantly strongly deleterious).
3. Quantification of Genomic Burden: Demonstrated that despite lower selection coefficients per site, non-coding regions contribute the majority of new deleterious mutations and beneficial substitutions in the genome due to their vast genomic footprint.
4. Improved BGS Modeling: Showed that incorporating non-coding DFEs significantly improves the prediction of genome-wide nucleotide diversity compared to models using coding regions alone.

4. Key Results

Inference Accuracy

• Site Number: Reliable inference requires >10,000 selected sites.
• Bias: Methods (DFE-alpha, GRAPES, fastDFE) accurately estimated parameters for moderately/strongly deleterious DFEs. However, for DFEs dominated by weakly deleterious mutations, the weak class was underestimated, and the moderate class was overestimated.
• Beneficial Substitutions ($\alpha$): Accurate estimation of $\alpha$ required ~100,000 sites. DFE-alpha and GRAPES performed best; fastDFE tended to overestimate $\alpha$ with smaller datasets.

Empirical DFE Estimates

• Coding Regions: Mutations were predominantly strongly deleterious ($2N_es_d \approx -700$ to $-2300$). The proportion of beneficial substitutions ($\alpha$) was high (~0.5).
• Non-Coding Regulatory Regions:
  • Mutations were primarily moderately deleterious ($2N_es_d \approx -40$ to $-100$).
  • The distribution was relatively even across deleterious classes (~25% in each class) for high-confidence regions.
  • Beneficial Substitutions: $\alpha$ was lower than in coding regions (0.19–0.52 depending on population and region) but still substantial.
• Specific Elements:
  • Promoters/Enhancers: Showed moderately deleterious DFEs.
  • TFBS: REDfly-validated TFBS showed a highly deleterious DFE (likely due to ascertainment bias toward strong phenotypes). In contrast, UniBind (computational) TFBS showed a moderately deleterious DFE similar to other non-coding regions.
  • Expression Correlation: TFBS near highly expressed genes experienced stronger purifying selection (more deleterious DFE) than those near lowly expressed genes.

Genomic Impact

• Deleterious Load: Non-coding regions (conserved non-coding + UTRs) account for ~87.7% of new deleterious mutations, while coding regions account for only ~12.3%.
• Beneficial Substitutions: Non-coding regions contribute ~69.7% of beneficial substitutions, compared to ~16% from coding regions.
• Background Selection (BGS):
  • Coding-only models predicted a mean $B$ (relative diversity) of 0.84.
  • Models including non-coding DFEs predicted a mean $B$ of 0.64.
  • Including non-coding regions improved the correlation between predicted and observed diversity (Pearson $R$ increased from 0.49 to 0.76).

5. Significance

This study fundamentally shifts the understanding of evolutionary forces in D. melanogaster by demonstrating that non-coding regulatory regions are the primary drivers of both deleterious load and adaptive evolution in terms of absolute numbers, despite having weaker selection coefficients per site than coding regions.

• Methodological: It highlights the critical need for realistic genomic architecture in simulations and the limitations of current DFE inference methods for weakly selected sites.
• Evolutionary: It suggests that ignoring non-coding selection leads to significant underestimation of background selection, resulting in inaccurate demographic inferences and misinterpretation of genomic diversity patterns.
• Biological: The finding that regulatory mutations are mostly moderately deleterious implies they persist longer in populations, exerting sustained selective pressure on linked neutral sites, particularly in low-recombination regions.

The authors conclude that accurate modeling of genome-wide variation and evolutionary history requires the explicit inclusion of non-coding DFEs.