Mach-1

Mach-1 is a long-context RNA foundation model for predicting transcriptome architecture. This repository houses the core model weights alongside the scripts needed to tokenize sequences, train and fine-tune the StripedHyena-based architecture, and run inference workflows.

Repository Structure

mach-1
├── processing-seqs/          # Tokenization configs and CLI tooling for data preparation
│   ├── mach_tokenizer.json    # Tokenizer configuration used across notebooks and scripts
│   ├── prepare_data.R         # RNA sequence preprocessing and formatting utilities
│   └── tokenize_data.py       # Batch tokenizer for genomic fastas/CSVs
├── training-model/           # Configuration, training, and inference scripts
│   ├── configuration_mach.py # Default StripedHyena model definition
│   ├── generate_seqs.py       # Synthetic sequence generation entry point
│   ├── get_embeddings.py      # Embedding extraction for downstream analyses
│   ├── get_likelihoods.py     # Likelihood computation and scoring helpers
│   ├── mach_dependencies.sh   # Environment bootstrap script
│   ├── modeling_mach.py      # Core Hyena architecture implementation
│   └── train_model.py         # Training script for Mach-1 checkpoints
└── model/                    # Pretrained checkpoints and tokenizer artifacts

Getting Started

1. Install dependencies listed in training-model/mach_dependencies.sh or adapt them to your compute environment.
2. Use the scripts in processing-seqs/ to prepare and tokenize the RNA sequences of interest.
3. Train or fine-tune Mach-1 with training-model/train_model.py, or run inference with get_likelihoods.py, get_embeddings.py, and generate_seqs.py.
4. Transfer the resulting outputs (likelihoods, embeddings, variant scores, synthetic sequences) into the directory structure expected by mach-1-manuscript to reproduce the manuscript analyses.

Companion Repository

The full set of data-processing pipelines, downstream analyses, and figure-generation workflows that accompany the Mach-1 study live in the companion repository mach-1-manuscript.