RegEvol: detection of directional selection in regulatory sequences through phenotypic predictions and phenotype-to-fitness functions

RegEvol is a novel computational framework that detects directional selection in regulatory DNA by integrating machine learning-based predictions of transcription factor binding with evolutionary models to infer fitness functions, successfully identifying adaptive signals in *Drosophila* and human genomes that were previously difficult to detect using sequence conservation alone.

The Gist

Imagine your DNA is a massive, ancient library of instruction manuals. For a long time, scientists have been very good at reading the "protein-coding" chapters of these manuals because they are like clear, rigid sentences. If a word changes in a sentence, it's easy to tell if the meaning is broken (bad) or improved (good).

However, the rest of the library—the regulatory DNA—is like the messy, handwritten sticky notes in the margins. These notes tell the cell when and where to read the main chapters. They are crucial for evolution (making a bat a bat and a human a human), but they are incredibly hard to study. Why? Because the "words" in these notes are fuzzy, and we didn't have a good way to tell if a change in a sticky note was just random scribbling or a deliberate edit by evolution.

Enter RegEvol, a new tool developed by Alexandre Laverré and his team. Think of RegEvol as a super-smart detective that can read these messy sticky notes and figure out if they were changed by accident or on purpose.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Conservation" Trap

Previously, scientists tried to find important changes by looking for conservation.

• The Old Way: Imagine you have a copy of a book from 100 years ago and one from today. If a sentence is exactly the same in both, you assume it's important. If it changed, you assume it didn't matter.
• The Flaw: This doesn't work well for sticky notes. Sometimes, the meaning of a note changes completely (evolution is happening!), but the letters look totally different. Or, a note might change slightly just by random chance (drift), and the old method would think it was important. It's like judging a recipe by looking only at the font size, not the ingredients.

2. The Solution: The "Function First" Detective

RegEvol changes the game. Instead of just counting how many letters changed, it asks: "What did this change do?"

• Step 1: The Crystal Ball (Machine Learning): The team trained a computer model (using something called a gkm-SVM) to act like a crystal ball. This model knows exactly how a specific change in a DNA letter affects a "molecular switch" (called a Transcription Factor).
  • Analogy: Imagine a dimmer switch for a light. RegEvol can predict: "If you turn this specific screw one click to the left, the light gets 10% brighter. If you turn it right, it gets 20% dimmer."
• Step 2: The Map of Possibilities: For every piece of DNA, RegEvol creates a map of all possible changes. It knows what happens if you change every single letter. This is the "Genotype-to-Phenotype" map.
• Step 3: The Detective Work (Evolutionary Models): Now, the detective looks at the actual changes that happened in a specific species over time. It compares the real changes to the map of possibilities.
  • Scenario A (Random Drift): The changes look like a random walk. Some went up, some went down, no pattern.
  • Scenario B (Stabilizing Selection): The changes try to stay in the middle. If a mutation makes the light too bright or too dim, it gets "erased" by nature. The system stays the same.
  • Scenario C (Directional Selection): This is the smoking gun! The changes are all pushing in the same direction. Maybe every change made the light slightly brighter. This suggests evolution was actively trying to make the light brighter.

3. Why This is a Big Deal

The paper tested RegEvol on millions of DNA regions in fruit flies and humans.

• In Fruit Flies: They found that about 5% of the regulatory switches were being actively "tuned" by evolution (Directional Selection). These weren't random changes; they were concentrated around genes for reproduction and immunity.
  • Metaphor: It's like finding that the only parts of a car being upgraded in a factory are the engine and the brakes. You know those are the parts that matter most for survival and speed.
• In Humans: Because human evolution is slower (fewer changes per generation), looking at one gene at a time was too hard to see the signal. So, the team used a "grouping" strategy. They looked at all the switches in the nervous system and male reproductive system together.
  • Result: They found a strong signal of active evolution in the brain and reproductive systems. It's like realizing that while one brick in a wall might look random, if you look at the whole wall, you can see it's being built into a specific shape.

4. The "Bias" Fix

Older methods had a flaw: they were easily tricked by "loud" changes. If one mutation caused a huge effect, the old tools would scream "Evolution!" even if it was just a lucky accident.
RegEvol is more like a wise judge. It doesn't just listen to the loudest voice; it looks at the whole choir. It asks, "Is everyone singing the same song?" If the changes are consistent, it's evolution. If it's just one loud note, it's probably noise. This makes the results much more reliable.

The Bottom Line

RegEvol is a powerful new lens for looking at the "dark matter" of our genome. It moves us from asking "Did this letter change?" to asking "Did this change improve the organism?"

By linking the physical DNA sequence to the actual function of the cell, and then to the survival of the species, this tool helps us finally understand how the "sticky notes" in our DNA are being rewritten to help us adapt, survive, and evolve. It's not just about the letters anymore; it's about the story they tell.

Technical Summary

1. Problem Statement

Detecting natural selection in non-coding regulatory DNA (e.g., enhancers, promoters) remains a significant challenge in evolutionary genomics.

• Limitations of Current Methods: Traditional approaches rely on sequence conservation (e.g., PhastCons, PhyloP) or substitution rate shifts. These methods are indirect; they infer selection from patterns of conservation or acceleration without modeling the functional consequences of mutations.
• Confounding Factors: Rate-based methods are susceptible to non-adaptive processes like biased gene conversion, mutation rate variation, and demographic history. They also struggle with fast-evolving elements or recently gained regulatory sites that lack deep conservation.
• Previous ML Approaches: A prior machine learning method (Liu & Robinson-Rechavi, 2020) used gapped k-mer SVMs to predict transcription factor (TF) binding changes ($\Delta$SVM) and compared observed changes to a null distribution via permutation tests. However, this approach suffered from:
  • Ascertainment bias: High-affinity binding sites in ChIP-seq data inflated false-positive rates.
  • Instability: Accuracy declined with evolutionary divergence as the non-parametric null model became unreliable.
  • Lack of mechanistic modeling: It did not explicitly model the fitness landscape or distinguish between stabilizing and directional selection effectively.

2. Methodology: The RegEvol Framework

RegEvol bridges the gap between machine learning-based functional prediction and explicit population-genetic models. It operates through a "Genotype $\to$ Phenotype $\to$ Fitness" pipeline:

A. Genotype-to-Phenotype Mapping (Functional Prediction)

• Model: Uses gapped k-mer Support Vector Machines (gkm-SVM) trained on ChIP-seq data to predict TF binding affinity.
• In Silico Mutagenesis: For every regulatory peak, the model computes the predicted change in binding affinity ($\Delta$SVM) for all possible single-nucleotide mutations.
• Distribution of Phenotypic Effects (DPE): This generates a landscape describing the range of potential functional changes accessible from the ancestral sequence.

B. Phenotype-to-Fitness Mapping (Evolutionary Modeling)

RegEvol defines three nested evolutionary scenarios using Beta distributions to parameterize the fitness landscape ($W$) relative to the phenotypic effect ($\Delta$SVM):

1. Neutral Drift: A flat fitness landscape ($\alpha = \beta = 1$). Fixation probability is independent of $\Delta$SVM.
2. Stabilizing Selection: A symmetric landscape centered at $\Delta$SVM = 0 ($\alpha = \beta \neq 1$). Mutations deviating from the ancestral optimum are selected against.
3. Directional Selection: An asymmetric landscape ($\alpha \neq \beta$). The fitness peak shifts, favoring mutations that increase or decrease binding affinity.

C. Likelihood Inference

• Data: Observed substitutions are identified by comparing the reference sequence to an inferred ancestral state using whole-genome alignments.
• Calculation: The method calculates the likelihood of the observed substitutions under each of the three models. It integrates the DPE (mutational availability) with the specific fitness function to determine fixation probabilities.
• Selection: A Maximum Likelihood Estimation (MLE) approach selects the best-fitting regime. Likelihood Ratio Tests (LRT) are used to compare models, with False Discovery Rate (FDR) correction applied for multiple testing.
• Aggregation Strategy: For short evolutionary branches with few substitutions (low statistical power), the framework aggregates likelihood differences across sets of regulatory elements (e.g., by tissue) using a SUMSTAT-like approach to detect coordinated signals.

3. Key Contributions

• Mechanistic Integration: Unlike previous methods that rely on conservation or simple permutation, RegEvol explicitly models the fitness consequences of mutations based on predicted molecular phenotypes.
• Bias Correction: It corrects for ascertainment biases inherent in ChIP-seq data and previous permutation tests, particularly regarding extreme-effect substitutions.
• Robustness to Divergence: By modeling the full distribution of effects rather than relying on outliers, RegEvol remains conservative and accurate across varying levels of evolutionary divergence.
• Scalability: The framework is flexible, allowing the integration of future deep learning models (e.g., BPNet, DeepSEA) as the predictive engine, provided they can generate $\Delta$SVM-like scores.

4. Key Results

A. Simulation and Validation

• Performance: RegEvol outperformed the previous permutation test in sensitivity and specificity across all conditions.
• False Positives: It maintained a very low False Positive Rate (FPR $\approx 10^{-4}$) even under stabilizing selection or neutral drift, whereas the permutation test showed inflated FPRs with increasing divergence or extreme $\Delta$SVM values.
• Power: Detection power increased with the number of substitutions. RegEvol could detect directional selection with as few as 4 substitutions, whereas the permutation test required significantly more.
• Asymmetry: The method successfully handled the natural asymmetry of mutational effects (where most mutations are deleterious) without inflating false positives.

B. Application to Drosophila melanogaster

• Dataset: Analyzed ~2.8 million TF binding sites across 740 experiments.
• Findings: Identified 5.1% of peaks under directional selection.
• Biological Validation:
  • Peaks under directional selection showed a significantly higher substitution-to-SNP ratio, consistent with recent selective sweeps.
  • Genes associated with these peaks were enriched in reproductive (testis, ovary) and immune tissues, aligning with known targets of rapid adaptation.
  • A weak but significant correlation was found between purifying selection on coding sequences and the proportion of directional selection on associated regulatory peaks.

C. Application to Human CTCF Binding

• Challenge: Short evolutionary distances in mammals resulted in low substitution counts per peak, limiting single-peak detection.
• Solution: Applied the tissue-level aggregation strategy.
• Findings: Aggregated signals revealed significant enrichment of directional selection in the Nervous System and Male Reproductive System. This confirms previous hypotheses about accelerated evolution in these systems but provides a regulatory-specific mechanism.

5. Significance and Implications

• Paradigm Shift: RegEvol moves regulatory evolution analysis from "rate-based" (conservation) to "function-based" (phenotypic impact), analogous to how $dN/dS$ ratios analyze coding sequences.
• Adaptation Insights: It provides a powerful tool to identify specific regulatory elements driving adaptation, particularly in systems where sequence conservation is low but functional change is high (e.g., immune response, reproduction).
• Future-Proof: The framework is agnostic to the underlying predictive model, meaning it can evolve alongside advances in deep learning for genomics, ensuring long-term utility.
• Conservative Inference: By prioritizing consistency of directional shifts over the magnitude of single mutations, RegEvol offers a robust, conservative approach to detecting positive selection, reducing false discoveries in noisy genomic data.

In summary, RegEvol represents a major advancement in evolutionary genomics by mathematically linking machine learning predictions of molecular function to population genetic theory, enabling the precise detection of adaptive evolution in non-coding regulatory regions.