The problem
Genomic risk screening is dominated by tooling built for bioinformaticians: command-line pipelines, file formats nobody outside the lab reads, and turnaround measured in hours. The people who need the answer — clinicians, researchers running a first-pass triage — are not the people the tooling was designed for. We wanted to find out whether a well-engineered gradient-boosted model behind a plain web interface could return a defensible risk classification from a DNA sequence in under a second, without asking the user to learn a pipeline.
The approach
Feature engineering over raw sequence
Raw nucleotide sequences are not directly usable by a tabular model, so the first half of the work was representation. We built a deterministic feature extraction stage that turns each sequence into a fixed-width numeric vector, then held that stage constant so every later accuracy comparison measured the model rather than the preprocessing.
- k-mer frequency encoding across multiple k values to capture local motifs
- GC content, sequence length and composition-ratio features
- Deterministic, versioned preprocessing shared by training and inference
Model selection and validation
XGBoost was chosen over a deep sequence model deliberately: on this dataset size it trained in minutes rather than hours, produced per-feature importances a domain reviewer could interrogate, and gave up nothing measurable in accuracy. Validation used stratified splits with a test set untouched until the final evaluation.
- Stratified train/validation/test split preserving class balance
- Hyperparameter search over depth, learning rate and subsampling
- 97.37% accuracy on the held-out test set, with the confusion matrix reviewed rather than the headline number alone
- Feature importance surfaced in the UI so a prediction can be questioned
Serving it to non-specialists
A Flask API loads the serialised model once at process start and exposes a single prediction endpoint; a React dashboard handles sequence input, validation and result presentation. Both are deployed on Railway, which kept infrastructure work proportional to the size of the project.
- Model loaded once at boot, not per request
- Input validation and clear error states for malformed sequences
- Results presented with confidence, not as a bare label
The outcome
GeneRisk AI classifies a submitted DNA sequence at 97.37% accuracy on held-out test data and returns the result, with model confidence and contributing features, in about a second — through a browser, with no bioinformatics pipeline for the user to install or operate.
- 97.37% accuracy on a held-out test set the model never saw during tuning
- Sub-second typical prediction latency end to end
- Explainable output: contributing features shown alongside every prediction
- Deployed and reachable as a normal web application, not a research notebook
A public URL for this product is not currently published. The figures above come from HEILC's own build and evaluation records for the project.