PVC

Privacy-preserving variant calling.

This repository contains the core research implementation.

How it works

1. Prepare public graph, panel, k-mer, and haplotype-block structures.
2. Compute evidence from the client's read counts and coverage.
3. Score candidate haplotype pairs using plaintext computation or three-party secure computation.
4. Reconstruct the private protocol's outputs at the client and produce genotype calls.

Protocol	Private computation	Client receives
Light	Scores secret-shared, client-computed emissions	Candidate score vectors
Medium	Scores secret-shared emissions and selects the winner privately	Winning one-hot vector
Heavy	Computes emissions from shared count/coverage selectors, scores candidates, and selects the winner privately	Winning one-hot vector

Public panel structure, LD blocks, and candidate haplotypes remain public.

Code map

Path	Purpose
src/pvc/pangenome/	Graphs, variants, k-mers, and probability tables
src/pvc/index/	Index preparation, LD blocks, and genotyping plans
src/pvc/client/	Client-side k-mer processing and reconstruction helpers
src/pvc/genotype/	Plaintext genotyping and VCF output
src/pvc/genotype/private/	Light, Medium, Heavy, role separation, and profiling
src/crypto/CrypTen/	Original upstream CrypTen runtime, configuration, and license
src/pangenie/, src/gatk/	PanGenie and GATK integration code
tools/GenotypeConcordance/	Genotype comparison utility
tests/	Secret-sharing and role-separated protocol checks

Runtime requirements

The tested runtime uses Python 3.10, PyTorch 1.13.1, and NumPy < 2. The original Facebook Research CrypTen is included at commit 775868a02d6dac50774ce376a55b01fbd8bd85b6.

Install

python3.10 -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[test]'
pvc --help
python -m pytest

Run

pvc --mode heavy \
  --cereal /path/to/read_UniqueKmersMap.json \
  --graph /path/to/Graph.json \
  --panel-vcf /path/to/panel.vcf.gz \
  --blocks-file /path/to/panel.blocks.det \
  --mean-kmer-abundance 30 \
  --sample-name SAMPLE \
  --output results/calls.vcf

Choose plaintext, light, medium, or heavy with --mode; python -m pvc provides the same interface. Private modes run on CPU with three computing-party processes and a separate upstream trusted preprocessing server. The parent remains the client. PVC_THREADS_PER_ROLE controls PyTorch worker threads (default 1); PVC_MPC_TIMEOUT_SECONDS controls the local timeout (default 3600).

PLINK and bcftools are needed if --blocks-file is omitted. PanGenie, GATK, and Jellyfish are needed only for preprocessing/comparison workflows that use them. Native executables and reference datasets are supplied separately.

Run a chromosome comparison with Slurm

Use the same prepared chromosome inputs for all four modes. For chr21, supply its graph, read-count JSON, panel, blocks, and sample truth VCF. Install PVC in .venv as above, then save this as validate.sbatch in the repository root and replace the input paths:

#!/usr/bin/env bash
#SBATCH --job-name=pvc-chr21-check
#SBATCH --partition=cpu
#SBATCH --cpus-per-task=5
#SBATCH --mem=128G
#SBATCH --time=12:00:00
#SBATCH --array=0-3%2
#SBATCH --output=results/chr21/logs/%A_%a.out
#SBATCH --error=results/chr21/logs/%A_%a.err
set -euo pipefail
cd "$SLURM_SUBMIT_DIR"

modes=(plaintext light medium heavy)
mode="${modes[$SLURM_ARRAY_TASK_ID]}"
out="results/chr21/$SLURM_ARRAY_JOB_ID/$mode"
mkdir -p "$out"

export OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1
export NUMEXPR_NUM_THREADS=1 PVC_THREADS_PER_ROLE=1
export PVC_MPC_TIMEOUT_SECONDS=42600 PVC_OT_PAIRWISE_PRF_PADS=1
export CUDA_VISIBLE_DEVICES=""

.venv/bin/python -u -m pvc --mode "$mode" \
  --cereal /path/to/chr21/read_UniqueKmersMap.json \
  --graph /path/to/chr21/Graph.json \
  --panel-vcf /path/to/chr21/panel.vcf.gz \
  --blocks-file /path/to/chr21/plink_blocks.blocks.det \
  --mean-kmer-abundance 30 \
  --sample-name HG00438 \
  --vcf /path/to/chr21/truth.vcf.gz \
  --output "$out/calls.vcf" \
  > "$out/stdout.log" 2> "$out/stderr.log"

Submit from the repository root. Create the Slurm log directory before submitting:

mkdir -p results/chr21/logs
job_id=$(sbatch --parsable validate.sbatch)
job_id=${job_id%%;*}
squeue -j "$job_id"
sacct -j "$job_id" --format=JobID,State,ExitCode,Elapsed,MaxRSS

The array runs plaintext, Light, Medium, and Heavy, with at most two modes running concurrently. Adjust the partition, memory, and time limit for your cluster; 128 GB and 12 hours are resource requests, not measured requirements. Each job writes to a new directory under results/chr21/<job_id>/. This reruns genotyping from prepared inputs; it does not rebuild the graph or process raw reads.