ORILINX is a large language model built on DNABERT-2 to detect the location of replication origins across genomes.
For development or to use the latest unreleased version:
git clone --recursive https://github.com/Pfuderer/ORILINX.git
cd ORILINX
pip install -e .
orilinx fetch_modelsThis installs ORILINX in editable mode and creates a console script named orilinx that runs prediction. The fetch_models command downloads the required model weights from Hugging Face (~900 MB total) and only needs to be run once.
After installation, use the orilinx command to predict replication origins:
# Basic usage: analyse all chromosomes
orilinx --fasta_path genome.fa --output_dir results
# Analyse specific chromosomes only
orilinx --fasta_path genome.fa --output_dir results --sequence_names chr1,chr2,chr3
# Analyse a specific region (must be ≥2000 bp)
orilinx --fasta_path genome.fa --output_dir results --sequence_names chr8:1000000-5000000
# Also generate CSV output alongside bedGraph
orilinx --fasta_path genome.fa --output_dir results --output_csv--fasta_path PATH(required): Path to the reference FASTA file. An index file (.fai) must be present in the same directory. Create it with:samtools faidx genome.fa--output_dir PATH(required): Directory where results will be saved.
--sequence_names SEQUENCES(default: all): Which chromosomes or regions to analyse. Use comma-separated names (e.g.,chr1,chr2) or a comma-separated list of regions (e.g.,chr1:2000-6000).--output_csv: Generate CSV files in addition to bedGraph files.--stride INT(default: 1000): Stride in base pairs (bp) between consecutive windows.--score {logit,prob}(default: prob): Output score format (probability 0-1 or raw logit).--batch_size INT(default: 32): Number of windows per batch. Increase for faster analysis if memory allows; decrease if you hit CUDA out-of-memory.--num_workers INT(default: 8): Background processes for data loading. Set to 0 if experiencing issues.--max_N_frac FLOAT(default: 0.05): Skip regions with >5% unknown bases ("N"). Adjust tolerance as needed.--disable_flash: Use standard attention instead of flash attention (slower but compatible with more GPUs).--verbose: Show detailed runtime information (prints model paths, device and runtime settings).--no-progress: Hide progress bars.--doctor: Check that required model weights are present, then exit.
For a complete list of options, run: orilinx --help.
# Analyse human chromosome 8 with CSV output
orilinx --fasta_path hg38.fa --output_dir results --sequence_names chr8 --output_csv
# Analyse multiple chromosomes
orilinx --fasta_path hg38.fa --output_dir results --sequence_names chr1,chr2,chr3,chrX
# Analyse specific genomic region (1 Mb region on chr8)
orilinx --fasta_path hg38.fa --output_dir results --sequence_names chr8:10000000-11000000
# Larger batches for throughput (VRAM-dependent; reduce if you hit OOM)
orilinx --fasta_path hg38.fa --output_dir results --batch_size 64
# Verbose output and no progress bars (for CI/logging)
orilinx --fasta_path hg38.fa --output_dir results --verbose --no-progressPlease see the documentation for detailed usage instructions, visualisation, and an example workflow.
Pfuderer PL, Berkemeier F, Nassar J, Crisp A, Jaworski JJ, Moore J, Sale JE, Boemo MA. A genome language model for mapping DNA replication origins. bioRxiv 2026. [Read on bioRxiv]
It is under active development by the Boemo Group based in the Department of Pathology, University of Cambridge.
Should any bugs arise or if you have basic usage questions, please raise a GitHub issue. For more detailed discussions or collaborations, please Email Michael Boemo at mb915@cam.ac.uk.
The core of this work was funded by the Cancer Research UK Cambridge Centre.