DGRM — Parallel GRM for VNTR

DGRM computes Genetic Relationship Matrices (GRM) from VNTR scaled dosage data. It is optimized for large cohorts, multi-threaded execution, and produces GCTA-compatible outputs.

Highlights

• Multi-threaded core with automatic block-wise computation for large N
• Empirical per-variant standardization with low-variance filtering
• GCTA-compatible outputs: .grm.bin, .grm.id, .grm.summary, .grm.N.bin

Build

cd dgrm
cargo build --release
./target/release/dgrm --help

Usage

./target/release/dgrm \
  --input <FILE> \
  --output <PREFIX> \
  --threads <N> \
  [--keep <KEEP_FILE>]

Examples:

# Basic run
./target/release/dgrm --input chr1_vntr_scaled_dos.txt --output chr1_grm --threads 8

# Keep a subset of samples (FID IID or single-column file)
./target/release/dgrm --input data.txt --output out --threads 8 --keep keep.txt

Input format

Tab-delimited, no header:

sample1	0.23	1.45	0.78	...
sample2	1.12	0.89	1.23	...
sample3	0.45	2.01	0.67	...

• Column 1: sample ID
• Columns 2..M: VNTR scaled dosage per variant (not necessarily 0/1/2)
• Rows are samples

Method (concise)

• Empirical standardization per variant: for variant k, compute mean μ_k and sample std σ_k over finite values; set z[i,k] = (x[i,k] − μ_k)/σ_k; missing values become 0.
• Filter variants with low information: keep only σ_k ≥ 1e−8 and count ≥ 2; let m be the number of kept variants.
• Compute GRM: G = (Z · Zᵀ) / m. For very large N, G is built block-wise.

Outputs (GCTA-compatible)

• <prefix>.grm.bin: GRM values as float32 in row-major lower-triangular order (including diagonal)
• <prefix>.grm.N.bin: number of variants used (float32, constant m for all pairs)
• <prefix>.grm.id: sample IDs (two columns: FID, IID)
• <prefix>.grm.summary: summary statistics

Note: This tool writes .grm.bin in the standard row-major lower-triangular order commonly accepted by GCTA utilities.

Performance & memory

• Full GRM in memory uses ~ N² × 8 bytes (f64). The tool automatically switches to a block-wise algorithm for large N.
• Set threads via --threads. Throughput typically scales near physical core count.
• For extra speed, set RUST_LOG=info and consider tuning DGRM_MM_CHUNK (e.g., 256/512/1024).

Logging

RUST_LOG=info ./target/release/dgrm --input data.txt --output result --threads 8 | cat
RUST_LOG=debug ./target/release/dgrm --input data.txt --output result --threads 8 | cat

License

MIT