> For the complete documentation index, see [llms.txt](https://sxu99.gitbook.io/ssblazer/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://sxu99.gitbook.io/ssblazer/use-cases/train-a-new-model.md).

# Train a New Model

Beyond SSB predictions, SSBlazer's framework allows for training on different lesion types, such as double-strand breaks, by supplying a dataset specific to the desired lesion type.

In this section, we will describe how to train a new model from scratch based on SSBlazer. 

{% hint style="warning" %}
We **strongly recommend** utilizing GPU devices for model training. This approach significantly enhances computational efficiency, thereby accelerating the training process.
{% endhint %}

Before beginning, ensure that you have installed `bedtools`, a powerful toolset for genome arithmetic. This tool is essential for preparing your data in the required BED (Browser Extensible Data) file format.

You can install `bedtools` on most UNIX-like systems (including Linux and MacOS) using the following command:

```
conda install -c bioconda bedtools
```

## Data Preparation

Both training and validation data should be provided in the form of a `bed` file describing break sites. which describes the break sites. For instance, we have collected Double Strand Break (DSB) sites from the dataset [GSM4047457](https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSM4047457) and converted the provided `bigwig` file into `bed`file:

```tsv
chr1	11377	11378
chr1	11472	11473
chr1	13194	13195
chr1	13522	13523
chr1	14172	14173
chr1	14208	14209
chr1	16389	16390
chr1	16427	16428
......
```

{% hint style="info" %}
In the bed file, only the `chrom`, `chromStart`, and `chromEnd` columns are considered. SSBlazer requires that `chromEnd-chromStart=`1, which describes peak sites. 
{% endhint %}

## Generating Datasets

First, we need to generate positive and negative sequences:

```sh
break_count=1470465 # Number of DSBs 

# Gen positive fasta
bedtools slop -i DSB.bed -g ../genome/hg38/hg38.chrom.sizes  -b 125 > DSB_251.bed
bedtools getfasta -s -fi ../genome/hg38/hg38.fa -bed DSB_251.bed -fo DSB_251_pos.fasta

# Gen negative fasta
bedtools random -l 251 -n $num -g ../genome/hg38/hg38.chrom.sizes > _neg.bed
bedtools subtract -A -a _neg.bed -b ./DSB.bed > _neg_final.bed
bedtools getfasta -s -fi ../genome/hg38/hg38.fa -bed _neg_final.bed -fo DSB_251_neg.fasta
rm _neg.bed _neg_final.bed
```

Then, create the training and testing sets. By default, the sequences from Chromosome 1 are set aside as the testing set:

```
python make_dataset.py --neg DSB_251_neg.fasta -neg DSB_251_pos.fasta
```

This script will create `train.csv` and `test.csv`.

## Training

Now, you can train the model using these datasets:

```
python train_from_scratch.py --train train.csv --test test.csv
```

The model weights will be saved in the `./models` directory. After the model is trained, you can use it to predict the break sites on new data by loading model weights.