Skip to content

DietNetwork

Genetic ancestry prediction from SNP genotypes with HussinLab's published Diet Network, behind a command line, dietnet, that takes PLINK files in and gives predictions out.

Install Quickstart

Rochefort-Boulanger C*, Scicluna M*, Poujol R, Grenier J-C, Carrier PL, Lemieux S, Hussin JG. A Transparent and Generalizable Deep Learning Framework for Genomic Ancestry Prediction. AJHG (2026), doi:10.1016/j.ajhg.2026.06.003. Models: Zenodo doi:10.5281/zenodo.16943452.

The published models were trained on the 1000 Genomes Project (1KGP, WGS 30x, GRCh38): 2,322 samples in 24 populations, 229,986 SNPs, five folds times three seeds (78, 79, 80), so 15 models.

What it is

A Diet Network is a classifier whose first layer is too large to learn directly (one weight per SNP per hidden unit), so a second, auxiliary network predicts those weights from per-SNP features. In the published model the features are each SNP's genotype frequencies per population.

This package is HussinLab's code (HussinLab/DIETNETWORK) with a layer around it:
The model and training are upstream's, unmodified The scripts in Dietnet_train_and_evaluate/ are byte for byte HussinLab's, pinned to one commit and checked by the test suite. dietnet runs them, and runs their model class for prediction.
Biobank scale Cohorts are read straight from .bed, matched to a model once and memory-mapped; every model in an ensemble reads the same copy. UK Biobank (486,748 samples) runs through the 15 published models in about 13 minutes on one A100.
Safer matching New cohorts are matched to the model's SNPs by position and allele. A SNP whose dosage counts the other allele is flipped, never silently mirrored. Missing SNPs are scaled the published way, by D/(D−|C|).
Portable model packages Each trained (seed, fold) is a self-contained directory that records its preprocessing and the upstream commit that trained it.
Other SNP features dietnet build-embedding assembles the auxiliary network's input from several sources, including AlphaGenome variant effects, without changing upstream's model.

Correct means identical to upstream

Where this package does something upstream's scripts also do, the definition of correct is that the two give the same result. tests/parity/ checks it against a full upstream checkout at the pinned commit: the HDF5 dataset, the prediction scores, and the published models' scores through upstream's own inference script. A change that moves a parity result is a bug, never a tolerance to loosen. See parity with upstream.

Where to start

  • Install: from source with uv or pip, the extras, and how to check the upstream code is intact
  • Quickstart: the published models on your cohort
  • Training: your own models, from PLINK files and labels
  • Tutorial: the whole pipeline on the bundled 1KGP fixture, in under a minute
  • SNP features: other inputs for the auxiliary network, including AlphaGenome
  • Formats: every file the pipeline reads and writes
  • Parity with upstream: how exactness is tested
  • Scaling: SLURM jobs for biobank cohorts and the reproduction
  • Command line: every command and option
  • Python API: the modules behind the commands