m6aformer 0.1.2

Deep learning toolkit for m6A site prediction (CNN + lightweight Transformer).

M6AFormer

Deep learning toolkit for transcriptome-wide N6-methyladenosine (m6A) site prediction, built on a hybrid CNN + lightweight Transformer architecture.

---

Highlights

• One model, multiple variants. Four pretrained checkpoints covering both 201 bp and 801 bp windows, with random and DRACH-filtered negative sampling strategies.
• Three ways to use it. A clean Python API, a m6aformer CLI, and an optional local web UI, all driven by the same core M6AFormer class.
• Frozen, audited thresholds. Each checkpoint ships with a decision threshold selected on inner-validation F1 (overridable at predict time).
• Reproducible. MIT licensed, fully open-source; bundled weights are shipped inside the wheel.

Pretrained models

Name	Architecture	Window	Negatives	Default
all_801	CNN + Transformer	801 bp	random	yes
all_201	CNN + Transformer	201 bp	random
drach_801	CNN + Transformer	801 bp	DRACH
drach_201	CNN + Transformer	201 bp	DRACH

Each model ships with a frozen decision threshold selected on inner-validation F1, recorded in its threshold.json. Users can override the threshold at prediction time.

Installation

# Basic (CPU/GPU inference + CLI)
pip install m6aformer

# With local web UI
pip install "m6aformer[web]"

# Development install (from a clone)
git clone https://github.com/zhixinniu/M6AFormer.git
cd M6AFormer
pip install -e ".[dev,web]"

PyTorch is intentionally pinned only as torch>=2.0. Install the CUDA-matching wheel from https://pytorch.org/get-started/locally/ first if you want GPU support.

Quickstart

Python

from m6aformer import M6AFormer

model = M6AFormer.from_pretrained("all_801")           # default checkpoint
df = model.predict_fasta("genome.fa", motif="DRACH")   # scan all DRACH As
# df columns: chrom, position, strand, motif_5mer, prob, label, ...

CLI

# List all bundled pretrained checkpoints
m6aformer list-models

# Score a FASTA file (auto-finds candidate adenosines)
m6aformer predict --input genome.fa --output sites.tsv --model all_801

# Score one or more inline sequences (no file needed)
m6aformer predict --seq ACGT...A...ACGT --strand + --model all_801

Local web UI

The [web] extra installs FastAPI + uvicorn and registers the serve subcommand. The bundled single-page UI exposes the same prediction pipeline as the CLI.

pip install "m6aformer[web]"
m6aformer serve                       # http://127.0.0.1:8000

For remote use over SSH (recommended over --host 0.0.0.0):

# On your laptop, forward the port and keep the server bound to localhost:
ssh -L 8000:localhost:8000 user@server
# Then open http://localhost:8000 in your browser.

Repository layout

M6AFormer/
├── src/m6aformer/      # the pip package (only this ships in the wheel)
├── training/           # training scripts (GitHub only, import m6aformer)
├── data_prep/          # dataset preparation scripts (GitHub only)
├── examples/           # runnable usage examples
├── tests/              # unit + integration tests
├── docs/               # documentation source

See training/README.md for how to retrain from scratch, and data_prep/README.md for the dataset construction pipeline.

Citation

If you use M6AFormer in your research, please cite the project. A machine-readable citation is provided in CITATION.cff.

License

MIT © 2026 Zhixin Niu.