README.md

# TReconLM synthetic test sets

This dataset contains the synthetic test sets used to evaluate TReconLM, a transformer-based model for trace reconstruction of noisy DNA sequences (see [our paper](https://arxiv.org/abs/XXXX.XXXXX)).

The real-world datasets used for fine-tuning are available here:

- [Microsoft Clustered Nanopore Reads](https://github.com/microsoft/clustered-nanopore-reads-dataset)
- [Noisy DNA Dataset](https://figshare.com/s/cd611884b34a8c89f4b4)

The corresponding test sets used in the paper can be reproduced using the preprocessing scripts in our GitHub repository under [`data/`](https://github.com/MLI-lab/TReconLM/tree/main/data).

## Synthetic Dataset Generation

Synthetic datasets are generated using [`data_generation.py`](https://github.com/MLI-lab/TReconLM/blob/main/src/data_pkg/data_generation.py). Each test set is created by:

- Sampling a ground-truth sequence of length *L*
- Introducing insertions, deletions, and substitutions with rates sampled uniformly from *[0.01, 0.1]*
- Randomly selecting the number of noisy reads *N* between *2* and *10*


### Files Included

- **`ground_truth.txt`**
  Contains the original DNA sequences, one per line

- **`reads.txt`**
  Contains the noisy traces (corrupted copies of the ground-truth sequences)
  - Each line is a single read
  - Clusters are separated by: `===============================`
  - The *i*-th cluster corresponds to the *i*-th line in `ground_truth.txt`

- **`test_x.pt`**
  A PyTorch tensor containing tokenized and padded input sequences used as model input, formatted as: read1|read2|...|readN : ground_truth

## Usage

Instructions for running inference using these datasets and our pretrained models are provided in the [`trace_reconstruction.ipynb`](https://github.com/MLI-lab/TReconLM/blob/main/trace_reconstruction.ipynb) notebook in our GitHub repository.