LinkedNN

Neural network for extracting LD features from SNPs

• Installation
• Usage
• Vignette

---

Installation

Quick start

pip install linkedNN

To test the installation you can apply the pretrained model from the paper to predict from a simulated dataset:

$ git clone https://github.com/the-smith-lab/LinkedNN.git
$ linkedNN --wd LinkedNN/Example_data/ --seed 1 --predict
using saved model from epoch 438
	test indices 0 to 0 out of 1
target 0 MRAE (no-logged): 0.136
target 1 MRAE (no-logged): 0.259
target 2 MRAE (no-logged): 0.289

The LD layer by itself can be accessed using:

from linkedNN.models import ld_layer

GPU compatibility: The code should work out of the box on a CPU, but to train on GPUs you need to sync torch with the particular CUDA version on your computer:

mamba install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

GSL installation

There may be additional requirements depending on the specific platform. In particular, installation of GNU Scientific Library (GSL) is sometimes needed for running simulations with msprime. See the msprime documentation (tskit.dev/msprime/docs/stable/installation.html) for up-to-date instructions.

---

Usage

The following are explanations of command-line flags for linkedNN.

Preprocessing

The program trains on datasets simulated usins msprime or SLiM. Before training, simulated tree sequences are preprocessed to (i) add mutations, (ii) sample SNPs, and (iii) write binary files. This can be applied to individual simulations, a range of simulation ID's, or all simulations in the specified directory; toggle this using the --simid flag. The working directory for linkedNN must itself contain a folder with tree sequences called TreeSeqs/ and a separate folder with the corresponding targets called Targets/, the latter saved as ".npy" format.

Example preprocessing command:

linkedNN --preprocess \
         --wd <path> \
         --seed <int> \
         --num_snps <int> \
         --n <int> \
         --l <int> \
         --hold_out <int> \
         --simid <int>

• preprocess: runs the preprocessing pipeline.
• wd: path to output directory.
• seed: random number seed ($>0$). The random number seed determines the names of outputs,so it's important to use different seeds for different analyses.
• num_snps: fixed number of SNPs to extract; it is recommended to use the number in your empirical dataset.
• n: number of diploid individuals; it is recommended to use the n from your empirical dataset.
• l: chromosome length; it is recommented to use l from your empirical dataset.
• hold_out: number of simulations from the full set to hold out for testing.
• simid: (optional) either (i) an individual simulation ID, (ii) a comma-separated range of IDs; if excluded, all ID's are preprocessed.

Training

After preprocessing all simulations, linkedNN can train a model using:

linkedNN --train \
         --wd <path> \
         --seed <int> \
         --batch_size <int>

• train: runs the training pipeline.
• batch_size: the size of mini-batches

Testing

To predict on held-out test data, run:

linkedNN --predict \
         --wd <path> \
         --seed <int> \
         --batch_size <int>

Empirical applications

To predict from an empirical VCF: leave in rare alleles, subset for a particular chromosome, and run the below command.

linkedNN --predict \
         --wd <path> \
         --seed <int> \
         --batch_size <int> \
         --empirical <path>

• empirical: is the path and prefix for the vcf file (without ".vcf").

---

Vignette

Below is a complete, example workflow with LinkedNN to provide a sense what inputs and outputs to expect at each stage in the pipeline.

Simulating training data

LinkedNN expects tree sequences, so you can use whatever program produces this output, i.e., msprime, SLiM, tsinfer. For this vignette, we will run one hundred small simulations using a script provided in the GitHub repo. However, note that 50,000 simulations and hundreds of training epochs may be required to train successfully.

git clone https://github.com/the-smith-lab/LinkedNN.git
for i in {1..100}
do
    echo "simulation ID $i"
    python LinkedNN/Misc/sim_demog.py $i 500,1e3 1e2,1e3 1e2,1e3 tempdir/
done

Preprocess

linkedNN --preprocess --wd tempdir/ --seed 2 --num_snps 5000 --n 10 --l 1e8 --hold_out 25

Train

linkedNN --train --wd tempdir/ --seed 2 --batch_size 10 --max_epochs 10

The new max_epochs flag is used here to limit the number of training epochs (default=1000).

Test

linkedNN --predict --wd tempdir/ --seed 2 --batch_size 10

---

How to cite: Smith CC. LinkedNN: a neural model of linkage disequilibrium decay for recent effective population size inference. arXiv. 2026. 10.48550/arXiv.2602.13121.

Source code: github.com/the-smith-lab/LinkedNN