Contents
VARIANT-GNN is a hybrid AI system that predicts whether missense variants in the human genome are pathogenic or benign. As team XYRA3, we applied to TEKNOFEST 2026 AI in Health Competition and successfully advanced past the preliminary stage. In this post, I explain what we built and why we made those decisions.
The Problem: The VUS Crisis in Human Genomics#
The human genome contains millions of variants. In genetic testing, a large portion of these variants are reported as "VUS" (Variant of Uncertain Significance). Determining whether these are pathogenic or benign is critical but extremely difficult.
The competition specification added an extra constraint: genomic coordinates are completely hidden. Label lookups from external databases are forbidden. The model must work solely with anonymous variant profiles — both a constraint and an accurate reflection of real-world conditions.
Architecture: Why a Single Model Isn't Enough#
The first question: why an ensemble? The ablation study gave the answer. Each model was learning something different: the GNN learned relational structure, boosting models learned tabular features, and the DNN learned deep representations. The models' errors didn't overlap — this made the combination powerful.
Weights were determined via Nelder-Mead optimization. XGBoost and LightGBM carry heavier weights, GATv2GNN and DNN carry lighter weights.
GATv2Conv: Why Not GAT?#
There's a critical choice in graph attention mechanisms. In classic GAT, the attention score depends only on the source node — static attention. The target node's features are not considered.
GATv2 (Brody et al. 2021) accounts for both source and target — dynamic attention. In a genomic context, this matters: the same variant should be able to form different relationships with different neighbors.
The architecture has 3 GATv2Conv blocks, each with LayerNorm, LeakyReLU, Dropout, and skip connections.
Graph Topology: Coordinate-Free k-NN#
Since genomic coordinates are hidden, a positional graph can't be built. Solution: cosine similarity-based k-NN graph. Variants are connected by similarity in feature space — no coordinates needed, fully compliant with the specification.
A critical technical detail: we build separate graphs for train and validation sets. Sharing the same graph would create data leakage.
Multi-Step Leakage-Free Preprocessing#
Each step was fit only on the training fold — only transform was applied to validation/test. The pipeline consists of these steps:
- ColumnAligner — aligns anonymous columns
- ACMGProxyFeatures — rule-based biological feature derivation
- SimpleImputer — missing value imputation
- RobustScaler — outlier-robust normalization
- BiologicalEnrichment — adds BLOSUM62 + Grantham scores
- SMOTE — optional, default off
- Feature Selection — ANOVA-based
- AutoEncoder — latent representation appended
- Cosine k-NN Graph — PyG format graph construction
Reliability Layer#
Model predictions pass through three layers:
Isotonic Calibration — calibrates raw probabilities. Fit only on the calibration set, the test set was never seen.
MC Dropout — multiple forward passes for GNN, standard deviation as uncertainty measure. High uncertainty → expert review flag. Observation: uncertainty in failed predictions is notably higher than in correct predictions.
OOD Detector — Z-score + Mahalanobis + KDE combination. Flags variants far from the training distribution.
Panel Structure#
The competition has 4 different panels: a general dataset, hereditary cancer panel, phenylketonuria (PAH), and cystic fibrosis (CFTR). Separate threshold optimization was done for each panel. For small, high-risk panels like CFTR, minimizing missed pathogenic variants was the priority.
Jury Evaluation#
We successfully advanced past the PSR stage. Full score was achieved on the Data and Method section — confirming the validity of the leakage-free pipeline and coordinate-free graph design. Explainability and learning process narrative are being deepened for PDR: GNNExplainer visual outputs, waterfall plots, and LIME-SHAP overlap analysis are being added.
SWA and Reproducibility#
SWA (Stochastic Weight Averaging) was applied during a portion of the later epochs — averaging checkpoints to improve generalization performance.
All RNG sources were locked with a fixed seed. Per the specification, the jury can run the code at any time and reach the same results.
PDR deadline: June 29, 2026. TEKNOFEST Şanlıurfa finals: September 30 – October 4, 2026.
The repository is not public while the competition evaluation is ongoing. For questions about the architecture, get in touch.