Genetic Fitness Paradox: Evo2 Analysis of Repeat Expansion Diseases

A computational biology project exploring how Genomic Foundation Models (Evo2) perceive pathogenic repeat expansions.

Project Overview

Repeat Expansion Diseases (REDs) like Huntington's (HTT), Friedreich's Ataxia (FXN), and Fragile X Syndrome (FMR1) are caused by the expansion of simple sequence repeats (e.g., CAG, GAA, CGG). Clinically, these expansions are deleterious.

This project asks a simple question: Does a state-of-the-art Genomic Language Model (Evo2) perceive these disease-causing mutations as "less fit"?

Using NVIDIA's Evo2-40b model, we analyzed the local genomic "fitness" (log-probability) of these repeats at varying lengths.

Key Finding: The Evo2 Paradox

Counter-intuitively, Evo2 predicts that expanded repeats have higher genomic probability (higher "fitness") than Wild-Type sequences.

Across all three diseases analyzed, the model's log-probability score increased (became less negative) as the repeat length grew from normal to pathogenic ranges.

Disease	Gene	Repeat	Trend Observed
Huntington's	HTT	CAG	Score ↑ with Length
Friedreich's Ataxia	FXN	GAA	Score ↑ with Length
Fragile X	FMR1	CGG	Score ↑ with Length

FMR1 (Fragile X): Note the increasing trend in Evo2 scores.

FXN (Friedreich's Ataxia): Expansion leads to higher predicted fitness.

HTT (Huntington's): Similar paradoxical trend observed.

Understanding the Metrics

Mean Log-Probability (Raw Score): This represents the model's "confidence" in the sequence. Scale: Log-probabilities are negative numbers (e.g., -0.7). Interpretation: Values closer to 0 (less negative) indicate higher genomic likelihood or "fitness" according to the model. Evo2 sees these sequences as more "natural" or evolutionarily probable.

Delta Score (Δ vs Wild Type): The difference in score between the Expanded Variant and the healthy Wild Type (WT). Formula: Score(Variant) - Score(WT) Positive Δ: The variant is predicted to be more fit than the Wild Type. Negative Δ: The variant is predicted to be less fit (deleterious). The Paradox: In our analysis, we consistently observe Positive Δ for pathogenic expansions.

Interpretation

This suggests a divergence between "evolutionary likelihood" (what the model learns from the genome) and "clinical organismal fitness."

• Hypothesis 1: The model may favor repetitive structures due to their high frequency in the genome (repeats are common).
• Hypothesis 2: The model might be capturing the "stability" of the DNA sequence itself rather than the downstream deleterious protein/RNA effects.
• Hypothesis 3: Expansions might be evolutionarily favored mechanisms for rapid variation, even if they risk disease (the "tuning knob" hypothesis).

Analysis Pipeline

The analysis was performed using a custom Python pipeline:

1. Data Preparation:

   • Curated FASTA files containing Wild Type (WT), Pre-Mutation, and Full Mutation sequences for HTT, FXN, and FMR1.
   • Sequences were extracted with upstream/downstream genomic context.

2. Evo2 Scoring:

   • Model: Evo2-40b (via NVIDIA NIM API).
   • Metric: Mean Log-Probability of the sequence.
   • Method: Zero-shot scoring of the repeat region + downstream context.

3. Visualization:

   • We compared the Evo2 Score against Repeat Length to determine the slope of fitness change.

Repository Structure

Evo2_Expansion_Project/
├── data/
│   ├── raw/           # Input FASTA sequences (e.g., FXN_merged.fasta)
│   └── processed/     # Evo2 Score CSVs (e.g., FXN_merged_evo2_results.csv)
├── figures/           # Generated plots showing Score vs. Length
├── src/               # Python analysis scripts
├── notebooks/         # Jupyter Notebooks for interactive data exploration
└── README.md          # This file

Usage

Prerequisites

• Python 3.8+
• pandas, matplotlib, seaborn
• NVIDIA API Key (for running new analyses)

Running the Analysis

To reproduce the plots or analyze new sequences:

python src/Evo2_REDs_Analysis.py data/raw/FXN_merged.fasta --motif GAA

Results generated: February 2026