How I Built a DNA Mutation Predictor Using ESM-2 and XGBoost (Open Source) As a Highskewler :]

The Problem

When a single nucleotide changes in DNA (a missense mutation), it can alter a protein's function. Some changes cause disease. Most don't. ClinVar - the NIH's database of known variants - has ~2,800 classified missense mutations. I wanted to build a classifier that could predict pathogenicity from sequence alone.

The Stack

• ESM-2 (facebook/esm2_t30_150M_UR50D) - a 150M-parameter protein language model from Meta. It reads protein sequences and produces 640-dimensional embeddings that capture evolutionary and structural information.
• XGBoost - gradient boosted trees, trained on 40 selected features from the ESM-2 embeddings.
• scikit-learn - feature selection (SelectKBest) and scaling (StandardScaler).

No deep learning at inference time. The ESM-2 embeddings are computed once and fed into a lightweight XGBoost model. The result: 82.5% accuracy on a 2,792-sample ClinVar dataset.

How It Works

1. You input a protein symbol + mutation (e.g., BRCA1 A1708E)
2. The tool fetches the protein sequence from UniProt
3. ESM-2 produces a 640-dim embedding for the mutated sequence
4. XGBoost classifies: PATHOGENIC or BENIGN
5. You get a confidence score + a protein feature graph

Running It

git clone https://github.com/NOOBHEKER/dna-mutation-predictor.git
cd dna-mutation-predictor
pip install -r requirements.txt
python -m src.cli

Or double-click predict.bat on Windows - it sets up everything automatically.

What I Learned

• ESM-2 embeddings contain enough signal to classify pathogenicity without hand-crafted features
• Feature selection matters: 40 out of 640 dimensions outperformed the full embedding
• CPU inference is slow (~30s per prediction) but acceptable for research use
• The hardest part was cleaning ClinVar data, not training the model

Try It

The project is open source under MIT license. Clone it, break it, improve it.

GitHub: https://github.com/NOOBHEKER/dna-mutation-predictor

If it's useful to your research, I'd appreciate a star or a coffee.