regLM: Designing realistic regulatory DNA with autoregressive language models

This repository contains code used to perform all experiments reported in the paper "regLM: Designing realistic regulatory DNA with autoregressive language models" by Avantika Lal, Tommaso Biancalani, and Gokcen Eraslan, along with trained model weights and synthetic regulatory elements designed by various methods.

The folder structure is:

- experiments.tar.gz:

    - yeast_promoters: Notebooks and synthetic sequences related to the experiments on yeast promoter sequence generation.

    - human_enhancers: Notebooks and synthetic sequences related to the experiments on human enhancer sequence generation.

    - scripts: Python scripts and functions used in both experiments.

Synthetic regulatory elements generated by regLM are available at the following paths:

- yeast_promoters/synthetic_promoters/lm_filtered.csv

- human_enhancers/synthetic_enhancers/lm_filtered.csv

The trained models are described below:

- yeast_reglm_epoch=9-step=580648.ckpt : regLM model trained on yeast promoter sequences

- yeast_regression_matched_epoch=8-step=65331.ckpt : sequence-to-expression regression model for yeast promoters trained on the same data as the regLM model

- yeast_regression_separate_complex_epoch=1-step=30824.ckpt  and yeast_regression_separate_defined_epoch=8-step=80568.ckpt : sequence-to-expression regression models for yeast promoters trained on the separate data from the regLM model

- human_reglm_epoch=24-step=65304.ckpt : regLM model trained on human_enhancer sequences

- human_regression_matched_epoch=7-step=20968.ckpt : sequence-to-expression regression model for human_enhancers trained on the same data as the regLM model

- human_regression_separate_epoch=4-step=4635.ckpt : sequence-to-expression regression model for human_enhancers trained on the separate data from the regLM model

Code to train, load and test these models is available in the experimental folders.