Moleflow

A multithreaded MinKNOW Read Until simulator written in Crystal. Stream synthetic or replayed nanopore signals over gRPC, test unblock/stop-data clients, or generate signals offline. Includes an R10 DNA current model.

Build

Requires Crystal >= 1.22, Shards, a C++20 compiler, curl, tar, unzip and patch. Run all commands from the repository root. Install dependencies for your OS:

# Ubuntu
sudo apt-get install -y build-essential pkg-config libnghttp2-dev libssl-dev libyaml-dev curl unzip patch

# macOS (Apple Silicon; requires Xcode Command Line Tools)
brew install pkg-config libnghttp2 openssl@3 libyaml
export PKG_CONFIG_PATH="$(brew --prefix openssl@3)/lib/pkgconfig:$(brew --prefix libnghttp2)/lib/pkgconfig:$(brew --prefix libyaml)/lib/pkgconfig${PKG_CONFIG_PATH:+:$PKG_CONFIG_PATH}"
export LIBRARY_PATH="$(brew --prefix libnghttp2)/lib:$(brew --prefix openssl@3)/lib:$(brew --prefix libyaml)/lib${LIBRARY_PATH:+:$LIBRARY_PATH}"

Then build:

sh scripts/build.sh

This installs Crystal dependencies and the native POD5 library, then builds bin/moleflow and bin/moleflow-pod5. Native downloads support Linux x86_64, Linux ARM64 and macOS ARM64; see pod5.cr for other platforms. Python is needed only for tests.

Dependencies have no commit pins in shard.yml. shard.lock records resolved versions; use shards update to refresh them.

Run

./bin/moleflow --validate -c examples/r10.json
./bin/moleflow -c examples/r10.json

The manager listens on localhost:9502 and the sequencing position on localhost:10001, without TLS by default. The simulation starts when a client sets up get_live_reads. Restart the process for a new session; use Ctrl-C to stop it.

An example receiving/unblocking client is available:

crystal run examples/client.cr -- localhost:10001

For deterministic offline simulation:

./bin/moleflow --simulate -c examples/r10.json

Configure

Start with examples/r10.json. Paths are relative to the working directory. Common settings:

FASTA inputs generate synthetic signals. For raw little-endian int16 inputs, set replay: true on the sample. Without a current model, synthetic inputs use a toy model. See src/moleflow/config.cr for all settings and validation rules.

Record and Replay

Set "output": "output" in your configuration before running. The directory must not already exist. Moleflow writes config.json, terminal read records in reads.jsonl, and per-read ADC files in signal/. Both live and offline runs can record output; stop_further_data stops delivery, not recording.

sh scripts/pod5.sh convert output run.pod5
sh scripts/pod5.sh prepare real.pod5 replay --limit=1000
./bin/moleflow -c replay/replay.json

POD5 conversion is separate from the server and holds one read in memory. Replay preparation normalizes calibration and rejects clipping or mixed sample rates. Use new destination paths. FAST5 is not supported.

Test

After building:

python3 -m venv .venv
. .venv/bin/activate
python3 -m pip install -r requirements-test.txt
sh scripts/test.sh

The suite checks Crystal formatting, Ameba, specs, gRPC/TLS, slow consumers and POD5 interoperability with an independent Python client. GitHub Actions runs it on Ubuntu and macOS.

Limits and References

Synthetic signals use fixed translocation speed and Gaussian noise. Biological signal realism and Dorado basecalling accuracy are not validated; use real signal replay for downstream pipeline testing. Live interaction timing affects results, so a seed alone does not reproduce an interactive run.

• API support and protocol limits
• Architecture
• Validation results and boundaries
• Audit history

Acknowledgements

Moleflow was developed with Icarust as a reference and uses its R10 model and adapter signal. Thanks to its authors and contributors. Moleflow does not reproduce Icarust's behavior exactly; see NOTICE for provenance and details.

License

Moleflow is licensed under MPL-2.0. Bundled model provenance and third-party license and redistribution notices are listed in NOTICE.