This repo contains the official implementation for the paper Dirichlet diffusion score model for biological sequence generation published in ICML 2023.
Dirichlet Diffusion Score Model (DDSM) is a continuous-time diffusion framework designed specificaly for modeling discrete data such as biological sequences. We introduce a diffusion process defined in probability simplex space with stationary distribution being the Dirichlet distribution. This makes diffusion in continuous space natural for modeling discrete data. DDSM is the first approach for discrete data modeling with continuous-time stochastic differential equation (SDE) diffusion in probability simplex space.
We showed that DDSM is capable of solving Sudoku and designing promoter sequences according to transcription initiation signals.
The Jax version of the code will be published soon.
Please create a new conda or pip environment specifically for running DDSM.
DDSM requires Python packages PyTorch (>=1.0). You can follow PyTorch installation steps here.
If you plan to run promoter designer model, DDSM requires Selene (>=0.5.0). For smoother experience, we recommend to install Selene via the following commands:
git clone https://github.com/kathyxchen/selene.git
cd selene
git checkout custom_target_support
python setup.py build_ext --inplace
python setup.py install
It may be needed to install some packages prior installation of Selene via these commands but the process should be straightforward.
Input data for sudoku and promoter designer experiment as well as model weights with the best performance can be downloaded from Zenodo
An example notebook containing code for applying a toy model to binarized MNIST dataset is here.
Usage.md contains detailed information how to use other scripts provided in the repository.
Time dilation is a generally applicable technique (not just for DDSM) for improving diffusion sample quality and is very easy to implement. It can be easily applied to other SDE-based diffusion models as well. It simply involves adding a c factor to the reverse diffusion process (c>1).
Time dilation works by biasing sampling toward higher-density areas, which often correspond to better-quality samples. It is advisable to increase the number of reverse diffusion steps by c, but it is not always necessary.
Another useful trick is to introduce time dilation only in the later part of reverse diffusion sampling, since it will avoid biasing sampling globally (e.g. in MNIST generation task, sampling more ones because one is the most frequent digit in MNIST) and only bias sampling locally(e.g. better digit image quality)
The evaluation is based on comparing generated sequences and human genome promoter sequences (ground truth) on the test chromosomes. The metric SP-MSE is the MSE between the predicted promoter activity of generated sequences and human genome sequences (lower is better). Our model trained with DDSM outperforms models trained with other approaches:
Model | SP-MSE |
---|---|
DDSM (time dilation 4x) | 0.0334 |
DDSM (time dilation 2x) | 0.0348 |
DDSM (time dilation 1x) | 0.0363 |
D3PM-uniform / Multinomial Diffusion | 0.0375 |
Bit Diffusion (one-hot encoding) | 0.0395 |
Bit Diffusion (bit-encoding) | 0.0414 |
One can find more benchmarks on various datasets in the paper (see Publications)
DDSM is distributed under a BSD-3-Clause license. See the LICENSE file for details.
DDSM is developed in Zhou lab at UTSW.
- Pavel Avdeyev
- Chenlai Shi
- Yuhao Tan
- Kseniia Dudnyk
- Jian Zhou
Pavel Avdeyev, Chenlai Shi, Yuhao Tan, Kseniia Dudnyk and Jian Zhou. "Dirichlet diffusion score model for biological sequence generation".
@InProceedings{avdeyev2023dirichlet,
title = {{D}irichlet {D}iffusion {S}core {M}odel for biological sequence generation},
author = {Avdeyev, Pavel and Shi, Chenlai and Tan, Yuhao and Dudnyk, Kseniia and Zhou, Jian},
url = {https://arxiv.org/abs/2305.10699},
booktitle = {International Conference on Machine Learning},
year = {2023},
}
The preferred way of asking questions about DDSM is the discussions tab. Before posting a question, consider looking through the existing threads - it is possible that your question has already been answered. To report any bugs, please use the issues tracker.
In case you prefer personal communication, please contact Pavel at Pavel.Avdeev(at)UTSouthwestern.edu or Jian at Jian.Zhou(at)UTSouthwestern.edu.