NAVI (Neural Accelerated Variance-component Inference)

Neural Accelerated Variance-component Inference (NAVI, 나비) is a neural estimator for inferring variance components of complex traits in large scale biobanks.

NAVI trains a graph neural network using simulated data from succinct tree sequences.

Once trained, it can instantaneously infer variance components of thousands of traits within seconds on a single GPU.

Examples

True parameter versus NAVI estimate across 5000 replicates. It takes 51ms on a single Nvidia A40 GPU.

REML versus NAVI estimates across 1000 replicates. NAVI is as precise as REML.

Dependencies

NAVI's neural network uses a jax backend and is implemented in flax.

pip install jax flax

Training data is simulated from succinct tree sequences implemented in tskit.

pip install tskit